information to users - brown.edu · pdf file• t.c. warburton and g.e. karniadakis the...
TRANSCRIPT
INFORMATION TO USERS
This manuscript has been reproduced from the microfilm master. UMI
films the text directly from the original or copy submitted. Thus, some
thesis and dissertation copies are in typewriter face, while others may be
from any type o f computer printer.
The quality of this reproduction is dependent upon the quality of the
copy submitted. Broken or indistinct print, colored or poor quality
illustrations and photographs, print bleedthrough, substandard margins,
and improper alignment can adversely affect reproduction.
In the unlikely event that the author did not send UMI a complete
manuscript and there are missing pages, these will be noted. Also, if
unauthorized copyright material had to be removed, a note will indicate
the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by
sectioning the original, beginning at the upper left-hand comer and
continuing from left to right in equal sections with small overlaps. Each
original is also photographed in one exposure and is included in reduced
form at the back o f the book.
Photographs included in the original manuscript have been reproduced
xerographically in this copy. Higher quality 6” x 9” black and white
photographic prints are available for any photographs or illustrations
appearing in this copy for an additional charge. Contact UMI directly to
order.
UMIA Bell & Howell Information Company
300 North Zeeb Road, Ann Arbor MI 48106-1346 USA 313/761-4700 800/521-0600
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Spectral/^ Methods on Polymorphic
Multi-Domains: Algorithms and Applications
by
Tim othy Warburton
B.A., Oxford University, 1993 Sc.M., Brown University, 1994 M.A. Oxford University, 1998
Thesis
Submitted in partial fulfillment of the requirements for the Degree of Doctor of Philosophy
in the Division of Applied Mathematics at Brown University
May 1999
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
UMI Number: 9932499
UMI Microform 9932499 Copyright 1999, by UMI Company. All rights reserved.
This microform edition is protected against unauthorized copying under Title 17, United States Code.
UMI300 North Zeeb Road Ann Arbor, MI 48103
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
© Copyright
by
Timothy W arburton
1999
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
This dissertation by Tim othy Warburton is accepted in its present form by the Division of Applied Mathematics as satisfying the
dissertation requirement for the degree of Doctor of Philosophy
Date \ I f ^1 1 &______ G -*—"George Em Karniadakis, Director
Recommended to the Graduate Council
Date J - j - Z z k #
Date 7 ^ , 7 f ? S'
Approved by the Graduate Council
D a t e VWH 7 / ] .Sd-Peder J. Estrup Dean of the Graduate School and Research
ISUa SjlZ 'Spencer Sherwin, Reader
Chi-Wang Shu, Reader
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The V ita of Timothy Warburton
Bom 5 October, 1971, Ipswich, United Kingdom
E ducation
B.A., University of Oxford, 1993
Sc.M, Brown University, 1994
M.A., University of Oxford, 1998
P u blication s
• T.C. W arburton, S.J. Sherwin, and G.E. Karniadakis, Spectral Basis Functions
for 2D Hybrid hp Elements, Submitted to SIAM J. Num. Anal., (1997).
• T.C. W arburton, I. Lomtev, Y. Du, S.J. Sherwin and G.E. Karniadakis, Galerkin
and Discontinuous Galerkin Spectral/hp Methods, CMAME Spectral, Spectral
Element and hp Methods in CFD, 1998.
M eetin gs and C onferences
• G.E. Karniadakis, T.C. W arburton and I. Lomtev, Spectral/hp Element Meth
ods on Unstructured Meshes, Fourth U.S. National Congress on Computational
Mechanics. August 6-8, 1997, San Francisco, California.
• T.C. W arburton and G.E. Karnidakis, Discontinuous Galerkin Methods for
MHD, 1998 SIAM Annual Meeting Ninth SIAM Conference on Discrete Math-iii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
ematics.
• T.C. W arburton, I. Lomtev, R.M. Kirby and G.E. Karniadakis, A Discontinu
ous Galerkin Method for the Compressible Navier-Stokes Equations on Hybrid
Grids. Tenth Int. Conf. on Finite Elements in Fluids, January 5-8, 1998, Tuc
son, Arizona.
• S.J. Sherwin, T.C.E. Warburton and G.E. Karniadakis, Spectral/hp Methods for
Elliptic Problems on Hybrid Grids, Tenth Int. Conf. on Domain Decomposition
Methods, August 10-14, 1997 Boulder, Colorado.
• T.C. W arburton and G.E. Karniadakis, Spectral Simulations o f Flow Past a
Cylinder Close to a Free Surface, ASME paper FEDSM97-3389.
• T.C. W arburton and G.E. Karniadakis The Wake of an Oscillating Cylinder
Close to a Free Surface, APS 49th Annual Meeting of the Division of Fluid
Dynamics, Syracuse, NY 24-26 November 1996.
• T.C. W arburton, S.J. Sherwin and G.E. Karniadakis, Hierarchical refinement
using spectral/hp triangles and prisms, Sixth Int. Conf. on CFD, September
4-8, 1995, Lake Taho, Nevada.
• T.C. W arburton, S.J. Sherwin and G.E. Karniadakis, Unstructured hp/spectral
elements: Connectivity and optimal ordering, Int. Conf. on Comp. Engin.
Science, Ju ly 30-August 3, 1995, Mauna Lani, Hawaii.
iv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A cknowledgm ents
In a large scale project it is impossible to carry on a lone quest and work in glorious
isolation. I therefore owe a debt of thanks to many people who contributed ideas,
thoughts, technology and most valuably criticism.
Professor George Karniadakis has been a very understanding mentor, helping
with his keen physical insight and providing many good ideas to keep the A /s k T ocr
project growing and developing. In some ways it was Dr. Spencer Sherwin who
gave me the greatest initial boost by bringing me into the ways of simulation science
with the original M etzT o c r code, and I will always be thankful for his insistance on
structured thought and code.
When I moved from developing the incompressible package, Igor Lomtev provided
a great source of information and help. His methods are a credit to him and he has
made my progress speed along.
M e n T o c r would not have taken on a life and style of its own if it had not received
so much challenging attention from the CRUNCH group. Dr. James Trujillo and Dr.
Ali Beskok have both been incredibly helpful with their expertise in fluids and their
garbage detectors. Dr. Ma Xia and Constantions Evangelinos have both pushed
and tested the code to great limits, which I initially thought were beyond its scope.
Robert (Mike) Kirby recently joined CRUNCH, yet in the last nine months he has
helped so much with mesh technology that I cannot see how we ever managed without
his spirited assistance. I feel tha t we have all benefited from cooperation and I hope
J\f s k T ocr matures with the next generation of CRUNCH.v
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Dr. Chi-Wang Shu and Dr. Jan Hesthaven have been very helpful answering my
( sometimes silly ) questions with patience and u n d e rs ta n d in g . They and the rest of
the faculty of Applied M ath have been so accommodating that I would like to thank
them all.
The N e k T cscr code does not stand on its own. It has been boosted by the support
of the METIS domain decomposition package[l], Henry Tufo’s GS communications
package [2], NASA’s FELISA mesh generation package [3] and Dr Tim B arth’s SIM
PLEX mesh generation package [4]. I cannot stress the importance of these codes
enough, they have been vital for rapid development.
This work was funded under the following grants:
• AFOSR: F49620-97-1-0185
• DOE: DE-FG02-95ER25239
• ONR: N000014-95-1-0256
• NSF: CTS-9417520
vi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C ontents
1 In trod u ction 1
1.1 O b je c tiv e s ........................................................................................................ 7
1.2 O u tlin e ............................................................................................................... 8
2 T h e E lem en ts 10
2.1 The T riangle..................................................................................................... 11
2.2 T he Q uadrila teral........................................................................................... 13
2.3 T he T e tra h e d ro n ........................................................................................... 15
2.4 T he P y r a m id ................................................................................................. 17
2.5 The P r i s m ........................................................................................................ 18
2.6 The Hexahedron ........................................................................................... 20
2.7 Global A sse m b ly ........................................................................................... 20
2.7.1 Algorithm 1 ......................................................................................... 23
2.7.2 Algorithm 2 ......................................................................................... 25
2.7.3 Algorithm For Connecting Prisms And Tetrahedra ................ 27
3 B asis F unctions 29
vii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
3.1 One-Dimensional Orthogonal P o ly n o m ia ls ................................................. 29
3.2 Singular Sturm-Liouville Problems on a Simplex in d-Dimensions . . 31
3.3 Orthogonal (Modal) B a s e s ............................................................................. 34
3.3.1 The Orthogonal Triangle Basis ........................................................ 37
3.3.2 The Orthogonal Quadrilateral B asis .................................................. 38
3.3.3 The Orthogonal Prism B a s is .............................................................. 38
3.3.4 The Orthogonal Hexahedral B a s is ..................................................... 39
3.3.5 The Decay of Basis C oeffic ien ts....................................................... 40
3.4 Semi-Orthogonal B a se s ...................................................................................... 41
3.4.1 Modal b a s i s ........................................................................................... 42
3.4.2 Nodal and Mixed Bases .................................................................... 51
3.5 Comparison of the B a s e s .................................................................................. 56
4 Operators 62
4.1 Elemental Spatial O p e ra to rs ........................................................................... 62
4.1.1 Integration ........................................................................................... 63
4.1.2 Flux In teg ra ls ........................................................................................ 64
4.1.3 Differentiation....................................................................................... 65
4.1.4 Implementation and P ro g ra m m in g ................................................ 67
4.2 Global Projection O p e r a t o r ........................................................................... 71
4.2.1 Global Convergence in Skew E le m e n ts .......................................... 72
4.3 Convective O p e ra to r ........................................................................................ 74
4.3.1 Accuracy of the Convective O p e ra to r ............................................. 78
viii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4.3.2 Spectrum of the Galerkin 2D Convective O perator.................... 79
4.3.3 Three-Dimensional Galerkin Convective Operator Spectrums . 83
4.3.4 Spectrum of the Galerkin 3D Convective O perator.................... 85
4.4 Diffusion O p e r a to r ......................................................................................... 88
4.4.1 The Schur Complement M ethod of Inverting Global Operators 89
4.4.2 Optimal Numbering for the Schur Complement Boundary M atrix 90
4.4.3 Numerical Results ............................................................................. 92
4.4.4 Convergence for the Galerkin Helmholtz O p e ra to r ...................... 96
5 Incom pressible F low Sim ulations 100
5.1 Incompressible Navier-Stokes E q u a t io n ..................................................... 101
5.1.1 Form ulation .......................................................................................... 101
5.1.2 Summary of Scheme, Boundary Conditions and Implementation 101
5.1.3 Wannier F lo w ...................................................................................... 103
5.1.4 Kovasznay F lo w ................................................................................... 105
5.1.5 Cylinder F lo w ...................................................................................... 106
5.1.6 The R obo tuna ...................................................................................... 110
5.2 ALE Incompressible Navier-Stokes ........................................................... 115
5.2.1 Form ulation .......................................................................................... 118
5.2.2 Temporal D isc re tiza tio n ................................................................... 119
5.2.3 Rayleigh Sliding Plate P r o b le m ..................................................... 121
5.2.4 Design of a Micro-Pump via S im ulation ........................................ 124
5.2.5 Micro-Pump Geometric Specifications........................................... 125
ix
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.2.6 Efficiency A n a ly s is ............................................................................ 128
5.2.7 Numerical S im ulation......................................................................... 129
6 Incom pressible V iscous M agnetohydrodynam ics 138
6.1 F orm ula tion ...................................................................................................... 139
6.2 Numerical Scheme ......................................................................................... 141
6.3 Magnetic Pearson’s V o rte x ............................................................................ 142
6.4 Orszag-Tang V ortex......................................................................................... 145
6.5 Cylinder F lo w ................................................................................................... 147
7 C om pressible F low Sim ulations 152
7.1 Numerical F orm ulation ................................................................................... 153
7.1.1 Discontinuous Galerkin for D if fu s io n ............................................ 154
7.2 Compressible Navier-Stokes Simulations ................................................... 156
7.2.1 Convergence......................................................................................... 158
7.2.2 Cylinder F lo w ..................................... 158
7.2.3 Flow Past a Two-Dimensional NACA0012 A ir f o i l ....................... 164
7.2.4 Flow Past a Multi-Body A i r f o i l ..................................................... 165
7.2.5 Flow Past A NACA0012 W ing W ith E n d p la te s ........................ 171
8 C om pressible V iscous M agnetohydrodynam ics Sim ulations 174
8.1 F orm ula tion ..................................................................................................... 175
8.2 The V • B = 0 Constraint ........................................................................... 176
8.3 Implementation of the Euler Flux T e r m s .................................................. 178
x
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8.4 Implementation of the Viscous T e rm s ....................................................... 183
8.5 Summary of the Compressible MHD Code Im p lem en ta tio n ................ 183
8.6 Two-Dimensional M agnetohydrostatic E x a m p le ................................... 185
8.7 Three-Dimensional M agnetohydrostatic Example ................................. 186
8.8 Compressible Orszag-Tang V o rte x .............................................................. 190
8.9 Flow Past a C y lin d e r..................................................................................... 197
9 Sum m ary 200
A Continuous E xpansion B asis 204
A .l Jacobi P o lynom ials........................................................................................ 204
A.2 Modal B a s is ..................................................................................................... 205
A.2.1 Two-Dimensional E x p a n s io n s ........................................................ 205
A.2.2 Three-Dimensional Expansions .............................................. 210
A.3 Nodal B a s i s ..................................................................................................... 218
A.3.1 Two-Dimensional E x p a n s io n s ........................................................ 218
A.3.2 Three-Dimensional Expansions .............................................. 220
xi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
List o f Tables
4.1 G LL implies Gauss Lobatto Legendre which is the Gauss quadrature
for a constant weight function with both x = ±1 points endpoints
included. G R Ja,p implies Gauss Radau Jacobi quadrature with (a, /?)
weights and the endpoint x = — 1 included................................................. 64
4.2 Timings for the Schur complement obtained on a single node SGI R8000. 93
5.1 Simulation param eters for the Kovasznay flow........................................... 106
5.2 Simulation param eters for the incompressible cylinder flow simulation. 109
5.3 Variation of Strouhal frequency with Reynolds number for incompress
ible flow past a two-dimensional c y lin d e r ................................................. 109
5.4 Variation of Cd with Reynolds number for incompressible flow past a
two-dimensional cylinder ........................................................................... 110
5.5 Simulation parameters for the Rayleigh problem....................................... 122
5.6 Simulation parameters for the micro-pump simulation............................ 131
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.7 Volumetric flow rate per unit width W as a function of Reynolds num
ber for closed-position piston-wall gap of g = 0.025L. The value for
v ~ 1 x 10-5 for both air and water are used. T he mass flow rate is
~M = pQ ..................................................................................................................................
6.1 Param eters for the equations of incompressible MHD..............................
6.2 Simulation parameters for the incompressible MHD Pearson’s vortex
sim ulation...........................................................................................................
6.3 Simulation parameters for the incompressible MHD Orszag-Tang vor
tex simulation....................................................................................................
6.4 Simulation parameters for incompressible MHD flow past a cylinder. .
7.1 Simulation parameters for compressible flow past a cylinder..................
7.2 Comparison of Strouhal frequency of incompressible against compress
ible (Mach=0.5) flow past a cylinder............................................................
7.3 Comparison of Strouhal frequency, drag and lift coefficients from hy
brid A fsK 'Tccr with results from other methods. Re=100, M=0.7 . .
7.4 Simulation parameters for compressible flow past a NACA0012. . . .
7.5 Simulation parameters for compressible flow past a m ultibody airfoil.
7.6 Simulation parameters for compressible flow past a NACA0012 airfoil
with endplates....................................................................................................
8.1 Variables and parameters used in the equations of compressible MHD.
8.2 Simulation parameters for the compressible Orszag-Tang Vortex. . . .
xiii
133
140
144
146
147
161
161
162
165
168
171
177
191
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8.3 Simulation parameters for compressible flow past a cylinder with a
magnetic field. ............................................................................................... 197
xiv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
List o f F igures
1.1 Top: Unstructured prism mesh, Bottom: p-type convergence for the
Helmholtz problem with A = 1 and solution sin(y)sin(z). The face
geometry was downloaded from the web-site: http://www.3dcafe.com/ 3
1.2 Hybrid discretization around a multi-element airfoil. Only part of the
domain is shown................................................................................................ 6
1.3 Hierarchy of the M en T 'o cr code .............................................................. 9
2.1 The tensor coordinates of the triangle.................................................. 12
2.2 The tensor coordinates of the quadrilateral......................................... 14
2.3 The tensor coordinates of the tetrahedron........................................... 16
2.4 The tensor coordinates of the pyramid................................................ 18
2.5 The tensor coordinates of the prism..................................................... 19
2.6 The tensor coordinates of the hexahedron........................................... 21
2.7 Diagram of the local top and base vertex of the tetrahedron.................. 22
xv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
2.8 Setting up the required connectivity for the discretization of a box as
shown in a ). Vertex A is given as the first local top vertex as shown in
b). In c) vertex B is then given as the local base vertex and the local
base vertices from group one are aligned to satisfy connectivity and the
final element orientation is shown................................................................ 24
2.9 Connectivity summary. Solid arrows imply we can connect elements
together with no extra mesh refinement, dashed arrows mean that there
are some constraints on allowable configurations...................................... 28
3.1 Mode shapes for the triangle modal basis (T rH H ) with N = 5. . . . 44
3.2 Top: Geometric ordering for modes, Bottom: Stiffness m atrix (V&, V<f)j)
for the continuous basis for triangles (N = 15) with the geometric or
dering of the rows and columns.................................................................... 46
3.3 Top: Polynomial degree ordering for modes, Bottom: Stiffness matrix
(V0i, V(j>j) for the continuous basis for triangles (N = 15) with modes
ordered by polynomial order......................................................................... 47
3.4 Mode shapes for the quadrilateral modal basis (Q uH H ) with N = 5. 48
3.5 Mode shapes for the quadrilateral mixed modal basis Q u H N with N = 5. 50
3.6 Mass matrix for the quadrilateral mixed modal basis (Q uH N ) with
N = 15............................................................................................................... 51
3.7 Mode shapes for the quadrilateral nodal basis (Q uN N ) with N = 5. . 53
3.8 Mode shapes for the triangle mixed basis (T rN H ) with N = 5............. 57
xvi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
3.9 Comparison of the elemental mass m atrix for (a) TrHH, (b) TrNH
using collocation, and (c) TrN H using exact integration for N = 15. 58
3.10 Comparison of the elemental stiffness m atrix for (a) TrHH, and (b)
T rN H bases for N = 15.................................................................................. 59
4.1 A piece of Jsf e tiT o c r code demonstrating how a list of elements can
be constructed.................................................................................................... 69
4.2 A piece of A fen 'T 'o tr code demonstrating how operations can be made
on a list of elements.......................................................................................... 70
4.3 Meshes (A-H) are consisted of three quadrilaterals and two triangles
which are progressively skewed by shifting the interior vertex............... 72
4.4 Convergence in the L2 norm for modal projection of the function u =
sin(rcx)sin{'Ky) on meshes A — H ................................................................. 72
4.5 Convergence in the L2 norm for mixed projection of the function u =
sin(irx)sin(7ry) on meshes A — H ................................................................. 73
4.6 Illustration of local and global numbering for a domain containing one
quadrilateral and one triangular elements. Here the expansion order is
N = 3, and we only show the boundary modes.......................................... 76
4.7 Z-matrix m ap from global to local degrees of f re e d o m ......................... 77
4.8 Exponential accuracy is achieved for the wave equation with u =
sin(7rcos(xx)) as initial condition................................................................. 78
4.9 Spectral radius of the weak convective operator on a periodic domain,
N = 12................................................................................................................. 79
xvii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4.10 Growth of the spectral radius of the Galerkin convective operator with
expansion order.................................................................................................. 80
4.11 pu{6) is an exact fit for the spectral radius of the Galerkin convective
operator on a periodic domain tiled with regular quadrilaterals and
w ith N = 12....................................................................................................... 81
4.12 p&.(9) is a close fit for the spectral radius of the Galerkin convective
operator on a periodic domain tiled with triangles and with N = 12. . 81
4.13 Spectral radius of the Galerkin convective operator on a periodic do
main discretized with non-regular elements................................................ 82
4.14 Left: Exploded spectral element mesh, Right: Variation of the spectral
radius of the convective operator in a periodic box discretized with six
tetrahedra and fourth-order expansion......................................................... 86
4.15 Left: Exploded spectral element mesh, Right: Variation of the spectral
radius of the convective operator in a periodic box discretized with six
pyramids and fourth-order expansion........................................................... 87
4.16 Left: Exploded spectral element mesh, Right: Variation of the spectral
radius of the convective operator in a periodic box discretized with two
prisms and fourth-order expansion................................................................ 87
4.17 Left: Exploded spectral element mesh, Right: Variation of the spectral
radius of the convective operator in a periodic box discretized with
eight hexahedra and fourth-order expansion............................................... 88
xviii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4.18 Form of the global operator m atrix. Notice the sparsity of A and the
block nature of B and C.................................................................................. 90
4.19 Here we show the domain tha t we solved the Helmholtz equation in the
following forcing function, / = — (3.47T2) sin (7r.x)cos(jry) cos (0.27r.z)
and boundary conditions u = sin (ir.x)cos(7r.y) cos with pe
riodic boundary conditions at the spiral ends. We see a dramatic
bandwidth reduction because of the aspect ratio of the mesh and the
exponential decay in the error with increasing expansion order............. 92
4.20 The exponential decay in the error with increasing expansion order for
the Helmholtz equation solved on the spiral domain................................ 94
4.21 Convergence test for the H elm h o ltz equation using quadrilaterals and
triangles, with Dirichlet boundary conditions. The exact solution is
u = sin(irx)cos(iry) and forcing function / = —(A + 2ir2)sin(Trx)cos(iTy) . 95
4.22 Convergence test for the Helmholtz, using a triangle and a quadrilat
eral, with Dirichlet boundary conditions: u = sm (7rcos(7rr2))............... 96
4.23 Convergence for the Helmholtz problem on a mesh of hybrid elements.
Dirichlet boundary conditions u = sin (x)s in (y)sin (z) and A = 10000. 97
4.24 Convergence is guaranteed for the Helmholtz problem even on very
skewed tetrahedra. Dirichlet boundary conditions u = s in ( y ) s m ( ^ ) s m ( y )
and A = 1........................................................................................................... 99
xix
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.1 Discretized solution domain for the Wannier-Stokes flow using 68 ele
ments. Left: Graph of exponential spatial convergence, Top Right: Hy
brid spectral element mesh, Bottom Right: Streamlines of the steady
state Stokes flow............................................................................................... 104
5.2 Steady state solution for the Kovasznay flow a t Reynolds number Re =
40 using a discretization using K = 24 elements. Left: The spectral
element mesh used, Right: Steady sta te streamlines................................ 105
5.3 Convergence in the L00 and Hi norms as a function of expansion order
for the steady state Kovasznay flow a t a Reynolds number Re = 40. . 107
5.4 Top Left: Three-dimensional spectral element mesh used to solve Ko
vasznay flow on, Top Right: Iso-contours of streamwise component of
velocity, Bottom: Convergence to exact solution with increasing order. 108
5.5 Domain and mesh for the two-dimensional incompressible flow past a
circular cylinder simulations................................................................. I l l
5.6 Instantaneous contours of vorticity for incompressible flow past a two-
dimensional cylinder at Reynolds numbers Re = 50,100 and 150. . . . 112
5.7 Instantaneous contours of vorticity for incompressible flow past a two-
dimensional cylinder at Reynolds numbers Re = 200 and 250..... 113
5.8 Top: The faces of the tetrahedra a t the nose of the fish. Bottom: The
spline surface on the body of the Robotuna..................................... 116
5.9 Robotuna simulation. Re=1000 based on height of tuna. N=5. Iso
contours of pressure are shown. Flow is from left to right 117
xx
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.10 Top: Spectral element mesh. K th = 50. Lower Left: Time convergence
history for zero initial conditions N = 5. Lower Right: Convergence
to exact solution with increasing N .............................................................. 123
5.11 Sketch of the micro-pump operating between two micro-channel sys
tems. Inlet and exit valves open and close periodicly with maximum
gap of Qrnax = 0.125L and minimum gap of gmin = 0.025L..........................125
5.12 Top: Deflection of the membrane y(x, t) = asiD.(nx/L) sin(o;£) Bottom:
Loci of the valve tips y (t) = ± ta n h (4 c o s (u ;t) ) ........................................ 126
5.13 Spectral element mesh used for the discretization of the micro-pump
system a t ejection-stage (top, membrane is moving up). The discretiza
tion of the flow domain during suction stage is shown a t the bottom
(membrane is moving down). The bottom figure also shows elemental
discretization obtained by 7th order modal expa n s ions used in M eKl'ar.1Z0
5.14 Non-dimensional volumetric flow rate variation w ith in a period of the
micro-pump, as a function of Reynolds number, R e = 9l l (a /L = 1/10,
a /h = 1 /3 ).......................................................................................................... 131
xxi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5.15 Vorticity contours for Re = 30 simulation. Top figure at t u j = 0.28,
corresponds to the beginning of the suction stage. Start-up vortices
due to the motion of the inlet valve can be identified. Middle figure
is a t t u j = 0.72, corresponding to the end of the suction stage. A
vortex jet pair is visible in the pump cavity. Bottom figure t u j = 84,
corresponding to early ejection stage. Further evolution of the vortex
je t and the start-up vortex of the exit valve can be identified...... 134
5.16 Close up of the vorticity contours for Re = 30 simulation at the left
valve (meshes shown on right side). Top: t u j = 0.28, corresponds to
the beginning of the suction stage. Start-up vortices due to the motion
of the inlet valve can be identified. Middle: t u j = 0.72, corresponding
to the end of the suction stage. A Vortex jet pair is visible in the
pump cavity. Bottom: t u j = 84, corresponding to early ejection stage.
Further evolution of the vortex je t and the start-up vortex of the exit
valve can be identified............................................................................. 135
5.17 Three-dimensional micropump simulation Re = 3. Top: An instant
during the suction stage. Bottom: An instant during the ejection stage. 137
6.1 Incompressible magnetic Pearson’s vortex (t = 2, instantaneous fields).
Top: Periodic spectral element mesh, Bottom Left: Pressure field,
Bottom Right: Magnetic streamfunction............................................ 143
xxii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
6.2 Incompressible magnetic Pearson’s vortex. Convergence plot for L2
error in x-component of magnetic field at time t = 2 versus expansion
order.................................................................................................................... 144
6.3 Incompressible magnetic Pearson’s vortex. Time accuracy plot for the
simulation run with N = 10. Convergence plot for L2 error in x-
component of velocity a t time t = 1 versus time step A t ............................. 145
6.4 Incompressible Orszag-Tang vortex ( t= l, instantaneous fields). Top
Left: x component of velocity, Top Right: y component of velocity,
Middle: Pressure, Bottom Left: x component of magnetic field, Bottom
Right: y component of magnetic field.......................................................... 148
6.5 Incompressible Orszag-Tang vortex ( t= l, instantaneous fields). Top
Left: Velocity streamlines, Top Right: Magnetic streamlines, Middle
Left: Vorticity, Middle right: Divergence of velocity, Bottom Left: Curl
of magnetic field, Bottom Right: Divergence of magnetic field.................... 149
6.6 Incompressible flow past a cylinder with inflow magnetic fields. From
the top: (1) x component of the velocity field, (2) y component of the
velocity field, (3) x component of the magnetic field, (4) y component
of the magnetic f i e l d ..................................................................................... 150
6.7 Incompressible flow past a cylinder with inflow magnetic fields aligned
with inflow velocity. Top: Pressure field, Middle: Vorticity, Lower:
Stream function of the magnetic f i e ld ....................................................... 151
xxiii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
7.1 Density contours obtained on a hybrid grid for an inviscid M = 0.3 flow
(Top), on a triangle grid (Middle). The bottom plot shows exponential
convergence of the error for the unstructured (triangles) and the hybrid
(squares) grid...................................................................................................... 159
7.2 Inviscid M = 0.3 flow past a bump in a three-dimensional (periodic in
the spanwise direction) domain. From the top: (1) domain, (2) spectral
element mesh used in the convergence test K = 120, (3) Iso-Contours
of density and (4) convergence of entropy to zero with increasing order. 160
7.3 Instantaneous iso-contours for the simulation of compressible flow past
a cylinder {Re — 100, M = 0.5) From the top: (1) density, (2) pressure
field, (3) x component of the velocity field, (4) y component of the
velocity f ie ld ...................................................................................................... 163
7.4 Instantaneous iso-contours of vorticity for the simulation of compress
ible flow past a cylinder {Re = 100, M — 0 .5 ) ........................................... 164
7.5 Top: Mesh of full domain for simulation of compressible, Mach 0.5,
Re=10,000 flow past a NACA 0012 airfoil at zero angle of attack to
the m ainstream flow, Middle: Mesh around body and wake, Bottom
Left: Close up of the airfoil, Bottom Right: Close up of part of the
wake region......................................................................................................... 166
xxiv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
7.6 The wake region (from f = 2 to f = 5) of a NACA 0012 airfoil a t
zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000,
3307 triangles, 4180 quadrilaterals, N = ll . Instantaneous iso-contours
of the density are shown................................................................................... 167
7.7 The wake region (from jr = 2 to £ = 5) of a NACA 0012 airfoil at
zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000,
3307 triangles, 4180 quadrilaterals, N = ll . Instantaneous iso-contours
of the divergence of momentum are shown.................................................. 167
7.8 The wake region (from f = 2 to f = 5) of a NACA 0012 airfoil a t
zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000,
3307 triangles, 4180 quadrilaterals, N = ll . Instantaneous iso-contours
of the curl of momentum are shown.............................................................. 167
7.9 Hybrid Mesh and close ups for the simulation of compressible flow past
a multi-body wing ......................................................................................... 169
7.10 Iso-Mach contours and streamlines for M = 0.5 flow past a two-
dimensional multi-component wing............................................................... 170
7.11 Skeleton mesh for flow past a three-dimensional NACA0012 airfoil with
endplates.............................................................................................................. 172
7.12 Iso-contours for x-component of momentum for M = 0.5 flow past a
three-dimensional NACA0012 airfoil with endplates.................................. 173
xxv
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8.1 Magnetohydrostatic test case for the compressible code. Top Left:
Mesh (K rri = 38, K qU(11i = 22) Top Right: Magnetic streamlines of
steady solution at N — 12, Bottom: dependence of steady state error
on expansion order........................................................................................... 187
8.2 Three-dimensional magnetohydrostatic test case for the compressible
code. Iso-contours of the magnetic and energy fields are shown a t time
t = 1. Top Left: Tetrahedral mesh used, Top Right: x-component of
the magnetic field, Middle Left: y-component of the magnetic field,
Middle Right: z-component of the magnetic field, Bottom: Conver
gence plot showing exponential decrease in L 2 error with increasing
expansion order................................................................................................. 189
8.3 Mesh used for the compressible Orszag-Tang vortex simulations on a
structured mesh................................................................................................ 191
8.4 Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.5).
Top left: density, Top right: energy, Middle left: x-component of mo
mentum, Middle right: y-component of momentum, Bottom left: x
component of magnetic field, Bottom right: y component of magnetic
field...................................................................................................................... 192
xxvi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8.5 Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.5).
Top left: curl of momentum, Top right: divergence of momentum, Mid
dle left: curl of magnetic field, Middle right: divergence of magnetic
field, Bottom left: momentum streamlines, Bottom right: magnetic
field streamlines................................................................................................. 193
8.6 Compressible Orszag-Tang vortex triangle mesh with K=132............... 194
8.7 Compressible Orszag-Tang Vortex ( t= l, instantaneous fields, Mach=0.2,
K=132). Top Left: Curl of momentum along the diagonal(N=4), Bot-
ton Left: Iso-contours of curl of momentum(N=4), Top Right: Curl of
momentum along the diagonal(N=6), Bottom Right: Iso-contours of
curl of m om entum (N = 6)............................................................................... 195
8.8 Compressible Orszag-Tang Vortex ( t= l, instantaneous fields, Mach=0.2,
K=132). Top Left: Curl of momentum along the diagonal(N=10), Bot-
ton Left: Iso-contours of curl of momentum(N=10), Top Right: Curl
of momentum along the diagonal(N=16), Bottom Right: Iso-contours
of curl of m om entum (N = 16)........................................................................ 196
8.9 Instantaneous iso-contours of the simulation fields for flow past a cylin
der (Sv = 100, Sr = 100, A = 0.1) with a magnetic field. From the top:
(1) density, (2) pressure, (3) x component of the velocity field, (4) y
component of the velocity field ................................................................. 198
xxvii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
8.10 Instantaneous iso-contours of the simulation fields for flow past a cylin
der (Sv = 100, Sr = 100, A = 0.1) with a magnetic field. Top: x com
ponent of the magnetic field, Bottom: y component of the magnetic
f i e l d ................................................................................................................... 199
A .l Vertex labelling for the standard quadrilateral r e g io n .......................205
A.2 Mass m atrix for the continuous basis for quadrilaterals (N = 15). . . 207
A.3 Vertex labelling for the standard triangular re g io n ............................. 207
A.4 Mass m atrix for the continuous basis for triangles (N = 15)..............209
A.5 Vertex labelling for three-dimensional standard r e g io n s ................... 210
A.6 Mass m atrix for the continuous basis for hexahedra (N = 15)...........213
A.7 Mass m atrix for the continuous basis for prisms (N = 15)..................215
A.8 Mass m atrix for the continuous basis for tetrahedra (N = 15).......... 218
A.9 Vertex labelling for the (a) standard quadrilateral and (b) standard
Hexahedral reg ion ............................................................................................ 219
xxviii
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter 1
Introduction
In spectral m ethods the question as to what basis functions (f>n(x) to use to approxi
mate a smooth function f ( x ) « Z)£=o fn4>n{x) in a separable domain is relatively easily
answered. Typically, boundary conditions, fast transforms, and numerical quadrature
dictate this choice depending on the particular partial differential equation considered.
In multi-domain spectral methods which are used in complex-geometry computational
domains the choice of the subdomain shape is an im portant factor as well. One can
choose arbitrary polygons, for example, to accommodate geometric complexity but
such a choice may prevent the use of polynomial basis functions and lead to prohibitive
computational difficulty in maintaining interfacial continuity among subdomains.
For the hp version of the finite element method the choice of subdomain shape
and associated functions bases dictates the efficiency of the method. Starting with
the pioneering work of B.A. Szabo in the mid-seventies several versions of this ap
proach have been formulated for both solid mechanics as well as fluid dynamics in
1
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
2the founding papers of Babuska et al [5, 6] and by Patera et al Patera84, KBP85.
Similar formulations were developed to allow greater flexibility in handling geometric
complexities and local refinement requirements in the works of [9], [10], [11] and [12].
In figure 1.1 we demonstrate how a complex three-dimensional volume can be formed
using prisms. We solved the Helmholtz problem within this volume, and show that as
we increased the polynomial order of our representation we obtained an exponential
decrease in the approximation error.
The basis or shape functions employed in the aforementioned formulations typi
cally involve either Legendre or Chebyshev one-dimensional polynomials, and multi
dimensional expansions are constructed using tensor products for quadrilateral or
hexahedral elements [13], [14]. A basis directly associated with boundary and inte
rior elemental nodes usually lacks hierarchy and in addition a variable p-order per
element is not readily achievable [7]. However, a modal basis can be constructed that
consists of vertex modes, edge modes, and interior modes in two-dimensions that are
hierarchical and allow for a variable order within as well as along the boundary of the
element [13]. In addition to the flexibility in handling more easily nonuniform reso
lution requirements, hierarchical bases can lead to better conditioning of mass and
stiffness matrices [14], and thus fast iterative algorithms can be effectively employed
in the solution algorithms.
The need for constructing p-type bases in triangular domains became evident
from the beginning. A. Peano [15] constructed a hierarchical triangular basis based
on area (barycentric) coordinates by selecting as nodal variables high-order tangen-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
3
1 cr*
10" 10"
i "-iio"
iff7icr* riff* - ■11» ■ •1 ■ »■11 ■«* ■1 • ■111 ■
4 5 6Expansion-Order
1 » * » • 1 ■ ■ » « 1
Figure 1.1: Top: Unstructured prism mesh, Bottom: p-type convergence for the Helmholtz problem with A = 1 and solution sin (y )s in (z). The face geometry was downloaded from the web-site: http://www.3dcafe.com /
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
4tial derivatives along each side evaluated at the midside and imposing appropriate
continuity constraints for C° and C l elements. A variation of this construction was
later developed [13] th a t introduces Legendre polynomials to avoid round-off error for
high-order p-expansions. Both approaches however require special integration rules
which are quite complicated a t high polynomial order. Dubiner [16] first developed
an alternative hierarchical basis for triangular domains, which unlike Peano’s basis is
based on cartesian coordinates and preserves the tensor product property that leads
to sum factorization and low operation count, a crucial property for p-type finite
elements and spectral methods.
Dubiner’s basis was implemented in [20] using a Galerkin formulation of the
Navier-Stokes equations, and it was found to be competitive in cost with the nodal ba
sis on quadrilaterals employed in the spectral element m ethod [7]. This new modal ba
sis employs Jacobi polynomials of mixed weight to automatically accommodate exact
numerical integration using standard Gauss-Jacobi one-dimensional quadrature rules.
In particular, exploiting the tensor product property of the basis, multi-dimensional
integrals can be evaluated efficiently as a series of one-dimensional integrals; similarly
the cost of evaluating derivatives or squares of a function is m aintained at operation
count 0 ( N d+1) where N is the number of modes per direction as in the quadrilateral
or hexahedral elements and d is the dimension of the element.
Recent trends in mesh generation have emphasized the use of hybrid discretiza
tions consisting of triangular, quadrilateral, tetrahedral, prismatic, pyramidal and
hexahedral subdomains in order to provide greater flexibility in refinement and dere
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
finement procedures. For high-order discretizations this implies tha t suitable basis
functions have to be developed to conform to the subdomain shapes while a t the same
time preserve the good approximation properties and efficiency of the aforementioned
bases. In this thesis, we develop a unified description of such hybrid basis functions
following earlier developments in [16] and [21]. We also develop five types of basis
functions which are either modal or nodal or mixed and which may or may not be
hierarchical.
There has been recently an interesting debate as to what are the relative advan
tages of structured and unstructured grids, with a renewed interest in cartesian grids
for complex-geometry aerodynamic flows [22]. The current confusion stems from the
fact that most finite element and finite volume formulations in use today produce
solutions, which depend strongly on the quality of the mesh. Specifically, for highly
distorted grids convergence is questionable, and in most cases convergence rates are
typically less than second-order. To this end, some - relatively few - efforts have
addressed the development of hybrid grids, i.e. a mixture of structured and unstruc
tured grids in order to combine the merits of both discretizations in the context of
complex-geometry aerodynamic flows [24], [25], [26]. Such methods lead to more flex
ible geometric discretization and resolution placement, but they are still of low-order
accuracy, i.e. a t most second-order.
We also develop a new formulation for compressible Navier-Stokes and compress
ible magnetohydrodynamics solutions employing high-order spectral/hp element dis
cretization on mixed element grids. A discontinuous Galerkin formulation is devel
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
6oped both for the convective as well as the diffusion contributions that allows multi
domain representation with a discontinuous (i.e. globally L 2) trial basis. This dis
continuous basis is orthogonal, hierarchical, and maintains a tensor-product property
(even for non-orthogonal elements), a key property for the efficient implementation
of high-order methods. Due to this basis orthogonality the resulting mass m atrix is
diagonal, and thus the proposed method is matrix-free given that an explicit time-
stepping is involved. The conservativity property is maintained automatically by the
discontinuous Galerkin formulation, and monotonicity is controlled by varying the
order of the spectral expansion, around discontinuities.
Figure 1.2: Hybrid discretization around a multi-element airfoil. Only part of the domain is shown.
An example of a hybrid discretization is shown in figure 1.2 for simulation of
flow past a multi-element airfoil. This problem was studied by Barth [4] who used
unstructured disretizations, i.e. only triangles. There are two regions of structured
discretizations as shown in these two plots: First, around the four airfoils where
boundary layers are present and where all the vorticity is generated. Second, in
the wake region where one single family of square elements is embedded within the
unstructured discretization. This uniform meshing minimizes dispersion errors of the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
7traveling vortex street and facilitates maximum storage efficiency. We will revisit this
example in section 7.2.1 in order to present simulation results a t subsonic speeds.
1.1 Objectives
The goals of this thesis are:
• To show how polymorphic elements can be built and interfaced to enhance
efficiency of the unstructured spectral element method (USEME).
• To demonstrate how Object-Oriented Programming can be used with minimal
overhead to provide a flexible toolkit for the polymorphic USEME.
« To provide theoretical and numerical evidence that USEME retains good ap
proximation properties even when very distorted meshes are used.
• To implement a semi-implicit Galerkin USEME scheme for the equations of
incompressible Navier-Stokes and magnetohydrodynamics (MHD) in the com
puter code f s fe n T ’a r .
• To investigate the effectiveness of USEME applied to moving domain problems
using the ALE formulation.
• To implement a fully explicit discontinuous Galerkin m ethod (DGM) USEME
scheme for the equations of compressible Navier-Stokes and magnetohydrody
namics in the computer code A/’e/cT 'ocr .
• To demonstrate accuracy for the above methods.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
81.2 Outline
The thesis will be laid out in the following way. In Chapter 2 we describe the elements
that we build domains with. In Chapter 3 we will outline the polynomial basis
functions th a t we use to approximate functions defined on these elements. There we
will also show how these elements can be connected together so th a t their coordinate
sysems line up. Chapter 4 details how we build operators on the elemental level and
on the global level (i.e. over connected elements). In Chapters 5 and 6 we show
how these building blocks can be applied to solving the incompressible Navier-Stokes
equations and the equations of viscous magnetohydrodynamics. In Chapters 7 and
8 we show how the Discontinuous Galerkin Method, and its local operators can be
used to solve the compressible Navier-Stokes equations and the compressible viscous
Ma.gnetohydrodynamic equations.
We have begun the unification of the Galerkin and Discontinuous Galerkin meth
ods under the umbrella code M e n T 'cxr . An outline of the current capability of this
code is dem onstrated in figure 1.3. All of the tree-root applications use the same set
of code library routines and the same data structures.
In the applications chapters of this thesis we will discuss a number of benchmark
tests and simulations using N skTTa r . For each simulation we will provide a table
of paxameters in order to reduce the amount of repetition involved. These tests are
very im portant in determining th a t the code is correct and demonstrating that the
underlying methods are robust and accurate when implemented.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
9
NeKtccrI
Galerkin
2d 2.5d1
3d
Steady . . F Steady Steady , PDomain Domain Domain
IDiscontinuous Galerkin
1
2d 3d
Incompressible Navier-Stokes Compressible
Figure 1.3: Hierarchy of the M eK/Tcxr code
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter 2
The E lem ents
We consider a number of element types that can be used to discretize two- and three-
dimensional spaces. In the hp community there has been a tendency to choose only
one element type to discretize a given domain. This is due to the widely conceived
idea that using more than one type of element is prohibitively expensive in terms
of implementation complexity. It is believed th a t this introduces a large number of
conditional statem ents. To counter these problems we have taken advantage of the
object-oriented language C-i—h- Object-oriented programming allows us to replace a
set of conditional “if” statements with a look up table for each element type which is
more computationally efficient. We will discuss the implementation aspects in section
4.1.4.
We show th a t it can be advantageuous to use both triangles and quadrilaterals
together to solve a problem set in a two-dimensional domain. In three-dimensions
we demonstrate th a t it is beneficial to use a combination of tetrahedra, pyramids,
10
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
11prisms, and hexahedra.
The most natural description of an element is as a subset of Rd, where d is the
dimension of the problem. Each element is then m apped to a reference element of its
own type. Though we only allow invertible mappings, these mappings can be quite
complex and can be represented as polynomial mappings in terms of the reference
elements coordinate system. This is commonly referred to as an isoparametric map
ping. Finally, we map the element to a tensor coordinate system attached locally to
the reference element.
In this chapter we develop new algorithms th a t enable the connection of three-
dimensional elements (i.e. tetrahedra, prisms and hexahedra) that have constraints
on the ways they can be connected together. Specifically we provide algorithms th a t
have an operation count that is proportional to the number of elements in a mesh.
2.1 The Triangle
We first consider the triangular shaped two-dimensional element. We will use the
formulation of Sherwin [20] to describe the element.
The natural coordinate systems will be given in term s of the ordered-pair x = (x, y ).
The coordinate system of the reference elements will be given in terms of (r, s ) . Finally
the local tensor space for each element will be given in terms of (a, b) coordinates.
The reference triangle, shown in figure 2.1, is described as the set of points:
T r ref = {(r, s)\ - 1 < r, s < 1; r + s < 0}
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
12
0.8
0.6
0.4
>.
-0.4
•0.6
•0.8
v*■ I ■ 1
r 1 * ' -1 i i I ■ • i i t • i t > I i • »-1 -0.5 0 0.5
Figure 2.1: The tensor coordinates of the triangle.
It is mapped to a straight-sided physical triangle with the following mapping:
(r + s) , (1 + r) 2 (1 + s) 3x = — —— v + ------ -v 2 + ------- -v 3
where the v 1^ 2, and v 3 are the physical coordinates of the vertices of the triangle
labelled in counter-clockwise manner.
The Jacobian for this mapping is:
§ ^ 4 = l(v2 - v l ) x ( v 3 - v I)|d{r, s)
The tensor (square) element is the set of points:
Qfensor = { ( a , b ) \ - l < a , b < l }
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
The tensor element is mapped to the reference triangle by:13
r =_ (1 + a) (1 - b) - 1
s = b
The Jacobian for this m apping is:
d(r, s) _ ( 1 - 6 ) d{a, 6) 2
We notice tha t this mapping is singular at 6 = 1. We will demonstrate that this
singularity will not present any numerical problems in chapter 4.
2.2 The Quadrilateral
Next we describe the quadrilateral. The reference quadrilateral, shown in figure 2.2,
is mapped to a straight-sided quadrilateral by the following:
( i - d t l - s ) l (l + r)(l-s) 2 (l+r)(l + s) 3 (l-r)(l + s) 4X — V " T " ------------------ " “ V “P ” V V
Again v 1 is an arbitrary vertex of the quadrilateral and v 2,v3,v4 are the vertices
labeled counter-clockwise from this vertex. The Jacobian for this mapping is:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
14
0.8
0.6
0.4
0 2
0>•-02
- 0.4
-0.6
-0.8
-1
Figure 2.2: The tensor coordinates of the quadrilateral.
d{x, y) 11 (v3 — v 1) (v4 — v 2) (v2 — v 1) (v3 — v4) (v4 — v 1) (v2 — v 3) ,= 2 2 X 2 + r 2 X 2 +S 2 X 2 1
This Jacobian includes three terms. If the quadrilateral is rectangular then the last
two terms are zero and the first term is ju s t the ratio of the area of the quadrilateral
and the reference quadrilateral. If the quadrilateral is not rectangular then the last
two terms are proportional to the sine of the angles between the two sets of edges:
t = + 1 , - 1 and s = + 1 ,-1 . The reference quadrilateral and the tensor element
coincide so the mapping between (a, b) and (r, s) is the identity mapping. Thus,
f^fj- is identically one. When we compare the overall Jacobian of the mapping from
physical element to tensor element we see th a t the triangle Jacobian is dependent on
b whereas the quadrilateral Jacobian is both a function of a and b. The reference
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
: v* i1 1 1 » » » > ->- » 1 -t- > « > » t I . l . r I
*1 -0.5 0 0.5 1X
15quadrilateral is the set of points Quref = {(r, s)| — 1 < r ,s < 1} like the reference
element for the triangle.
We next consider a set of elements commonly used to discretize a three-dimensional
volume. There are now four elements: the tetrahedron, the pyramid, the prism and
the hexahedron. Again we will consider the physical, reference and tensor represen
tations for each element which were proposed in [27].
2.3 The Tetrahedron
The natural coordinate systems will be given in terms of the ordered triplet x =
(x ,y ,z ) . The coordinate system of the reference elements will be given in terms of
(r, s, £). Finally the local tensor space for each element will be given in terms of
(a, 6, c) coordinates.
The reference tetrahedron is described as the set of points:
Teref = {(r, s, £)| — 1 < r, s, t < 1; r + s -1-1 < 1}
The reference tetrahedron is mapped to a straight-sided physical tetrahedron with
the following mapping:
(1 t + s -f- £) . ( 1 + r ) 2 (1 + s) 3 (1 + £) 4X = ------------------V1 + ^ v + 2 + 2
where the v 1^ 2^ 3 and v 4 are the physical coordinates of the vertices of the tetahe-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
16dron.
The tensor element is the set of points:
QtensoT = {(a.6 .c)| - 1 < a ,b ,c < 1}
The tensor element is mapped to the reference tetrahedron by:
r =
s =
(1 + a ) (1 — b) 2 2
(1 + 6)( ! — c) — 1
t = c
We notice that this mapping is singular a t b = 1 and c = 1.
V'
Figure 2.3: The tensor coordinates of the tetrahedron.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
2.4 The Pyramid
The reference pyramid is described as the set of points:
17
Pyref = {(r, s , t ) | — 1 < r, s , t < 1; r + t < 0 ; s + t < 0 }
The pyramid will be useful for making bridges between tetrahedra and hexahedra or
between prisms and hexahedra. The reference pyram id is mapped to a straight-sided
physical pyramid by:
_ (r + t) (s + t) 2 x ^ (1 + r ) ( s + t) 2 22 2 (1 - t ) 2 2 ( 1 - t )
(1 + r) (1 + s) 2 3 _ (r + 1) (1 + s) 2 4 (1 + f) 52 2 ( 1 - t ) 2 2 ( 1 - t ) 2
The tensor element is mapped to the reference pyramid by:
r =
t = c
Clearly, this mapping has a first order singularity in the r, t and s , t coordinates
making it have a second-order singularity a t r = s = — 1, t = 1. However, this
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
18is only a singularity due to the choice of coordinates and this singularity will be
removed by the volume Jacobian in a volume integral. Also, we will use quadrature
for the pyramid that does not include the top vertex, hence avoiding the calculation
of derivatives at this point.
Figure 2.4: The tensor coordinates of the pyramid.
2.5 The Prism
The reference pyramid is described as the set of points:
Pfref = (O', S, £)| - 1 < r, s, t < 1; r + 1 < 0}
The reference prism is m apped to a straight-sided physical prism by:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
(r + t) (l-s)„ i , (l+ r)(l 2 2
(r + t) (1 4- s)2 -v«
«/<* a
ettj,ePtsJ a
a)3 " v * ( i
Is PinPpetf to3 v *i~ (1 +
tb e te&*)
eroPce Ptis ,'*bIK:
r - . f t +^ Y -Z
- aPlS ? i s sin
S t
e
’W e*£■
6e tejj,s0tC°°r<]l'*abes
P tiPerfr>is-
Stti.s/or7
0ftt)e c°pyir'9ht °Wn{'Or
er rePr(° c/,c%0'’'■“a*®eC/ U,:,with,outPormr.rP'Ssi,Sion.
202.6 The Hexahedron
The reference hexahedron is described as the set of points:
Heref = {(r, M ) | - 1 < r ,s , t < 1}
The reference hexahedron is mapped to a straight-sided physical hexahedron by:
_ (1 — r ) ( l - a) ( 1 , (1 + r ) (1 ~ s) (1 ~ t) __2
2 2 2 2 2 2 | (1 + r) (1 + 5) ( l - i )v 3 + (1 — r) (1 + s) ( l - * ) v 4
(1 ~ r ) ( 1 - 5 ) ( l + *)v s + (1 + r) ( 1 - s ) (l + ^)v a
(1 + r) (1 + s) (1 + 1) + (1 — r) (1 + s) (1 + 1) g
The tensor element is mapped to the reference hexahedron by the identity.
2.7 Global Assembly
The elements we have described will be tessellated together to construct a continuous
solution domain. We shall only permit elements to connect by sharing common
vertices, complete edges and/or complete faces. Such a connectivity is commonly
referred to as a set of conforming elements. In this section we shall discuss issues
related to the process of globally assembling the elemental bases described in chapter
3.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
21
V'
Figure 2.6: The tensor coordinates of the hexahedron.
When two elements share an edge it is im portant for them to be able to determine
if their local coordinate system a t the shared edge are aligned in the same direction.
For instance, the mode shapes used in the C° methods must be continuous across
element boundaries. If the local coordinate systems are not aligned in the same
direction then edges modes of odd polynomial order will have different signs and the
odd edge modes of one of the edges will need to be multiplied by —1.
This condition becomes more complicated in three-dimensions when two elements
share a face. In this case it is not automatic that their coordinate systems on the com
mon face will line up. Considering the triangular faces of the tetrahedron, pyramid
and prism we see th a t there is a vertex on each triangular face th a t the coordinate
system for that face radiates from. We will call this vertex the face origin as it is
similar to a polar coordinate origin. The alignment constraint necessitates tha t when
two triangular faces m eet their origin vertices must coincide. Initially, it is not obvi-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
22
v4 (local top vertex)
v
v3 (local base vertex)
Figure 2.7: Diagram of the local top and base vertex of the tetrahedron.
ous how to satisfy this constraint for a mesh consisting of just tetrahedral elements.
We outline two algorithms tha t will satisfy this constraint. The first is based on the
topology of the mesh. We will only use the connections between elements to deter
mine how we should orientate elements. In the second method we will assume that
each unique vertex in the mesh will have been given a number. This second method
works under some loose conditions but is extremely easy to implement and is very
local in its nature.
It is useful to observe that one of the vertices of a tetrahedron is the face origin
vertex for the three faces sharing tha t vertex. We will call this the local top vertex.
Then there is one more face origin vertex on the remaining face which we call the
local base vertex. These are shown in figure 2.7.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
232.7.1 A lgorith m 1
Given a conforming discretization we can generate the local orientation of the tetra-
hedra using the following algorithm. We assume th a t we have both a list of vertices
and a list of elements which touch each vertex. This list of elements will be called a
vertex group and gill elements are assumed to have a tag of zero.
For every vertex in the list:
• Orient all elements with a tag of one in this vertex group so that their local base
vertex points at this vertex. Then set their tags to two.
• Orient all elements with a tag of zero in this vertex group so that their local top
vertex points at this vertex. Then set their tags to one.
This algorithm visits all vertices in the mesh and if this is the first time the
elements in the vertex group have been visited the local top vertex is orientated to
this vertex. If this is the second time the elements in the vertex group have been
visited then set the local base vertex to this vertex. To see how this works we can
consider the example shown in figure 2.8.
Here we assume that we are given a discretization of a box using six tetrahedra
as shown in figure 2.8a. Starting our algorithm we begin with vertex A. Since all
elements have a tag of zero at this point we go straight to the second part of the
algorithm and orientate all elements that touch this vertex so that their local top
vertices point to A. Therefore tetrahedra HBDA and BHEA are orientated as shown
in figure 2.8b and now have a tag set to one. Continuing to the next vertex B we
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
24
DA.
c) Final d ement orientation a t cod o f algorithm
Figure 2.8: Setting up the required connectivity for the discretization of a box as shown in a). Vertex A is given as the first local top vertex as shown in b). In c) vertex B is then given as the local base vertex and the local base vertices from group one are aligned to satisfy connectivity and the final element orientation is shown.
see tha t all elements belong to this vertex group. The first part of the algorithm is
to orientate the elements with a tag of one to have their local base vertex pointing
at B. So the tetrahedra HBDA and BHEA are ro tated as shown in figure 2.8c and
their tags are set to two. The second part of the algorithm then orientates all the
other tetrahedra to have their local top vertex pointing a t B. The connectivity is
actually satisfied at this point since the orientation the faces have on the boundaries
is irrelevant. However, if we continue the algorithm looping through the vertices
consecutively we end up with the tetrahedra orientated as shown in figure 2.8c.
Clearly, the connectivity is not unique since any elements that have their local top
vertex pointing at E can be rotated about E. We have demonstrated however tha t it
is possible to satisfy the connectivity requirements imposed by the co-ordinate system
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
and thereby im ply tha t the requirement is non-restrictive.25
2.7.2 A lg o r ith m 2
Assuming th a t every global vertex has a unique number, then for every element we
have four vertices with unique global numbers:
• Place the local top vertex at the global vertex with the lowest global number.
• Place the local base vertex at the global vertex with the second lowest global
number
• O rientate the last two vertices to be consistent with the local rotation of the
element (typically anti-clockwise).
It has been stated before that since the coordinate systems on the faces of the
tetrahedra are not symmetric, it is too difficult to use these coordinate systems [19].
We have shown, however that it is possible in linear or even constant time to satisfy
this constraint for any given tetrahedral mesh. This algorithm is local to each element
and should be implemented at a pre-processing stage.
We now extend this approach to include meshes consisting of tetrahedra, prisms,
and hexahedra. Unfortunately, in this case we find counter-examples where it is not
possible to satisfy the origin alignment constraint. We have isolated the problematic
cases, and they are unlikely to come up when using a mesh generator.
First we deal with the case when a quadrilateral face is shared by two elements.
In this instance it is sufficient to simply make the coordinate directions agree by
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
26reversing either face coordinate if necessary.
We now investigate over-constrained meshes. These cases can occur when prisms
and tetrahedra are used together in a mesh. We will use these examples to motivate
the actual algorithm we propose. The cost of this algorithm also depends linearly on
the number of elements in the mesh.
It is instructive to construct a chain of prisms. This is simply a long prism,
with equilateral triangular faces, divided a t intervals along its length into a set of
prisms connected at their triangle faces. The connectivity constraint requires that
the coordinate origins of the triangular faces must meet at every prism-prism interface.
This condition enforces that the collapsed edge m ust run in a continuous line through
the one edge of the original long prism. Now we twist the chain around in a loop
and connect its triangle ends. The chain now forms a closed loop of prisms. The
orientation of the end faces of the original chain must also satisfy the connectivity
constraint when they meet. However, we can make the chain into a Mobius band by
twisting it around the axis along its length. In this case the connectivity cannot be
satisfied without changing the mesh.
We can construct a second counter-example, this time involving one tetrahedron
and two chains of prisms. We construct two chains of prisms as outlined above, and
we join the tetrahedra into the prism chain by connecting two of its faces to the prism
chain triangular end faces. We repeat this operation, again connecting the remaining
two faces of the tetrahedron to the end faces of the second chain of prisms. We can
now repeat the twisting of the prism loops. This over-constrains the tetrahedron so
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
27that it cannot be oriented to satisfy the connectivity condition.
These two cases indicate that we cannot allow prism chains to reconnect into
closed loops and still satisfy our constraints. Also, we should not allow a prism chain
to connect to a tetrahedron with more than one of its triangular faces. If we only
consider meshes tha t satisfy these two constraints, then the following algorithm will
satisfy the connectivity constraints:
2.7.3 A lg o rith m For C onnecting P rism s A nd T etrahedra
• Find all prism chains in the mesh.
• Create a virtual connection between the faces of the tetrahedra that meet the
triangular faces at each end of the chain.
• Proceed with Algorithm 2 to connect the tetrahedral mesh treating the virtual
links as real connections.
• Orient each prism in each chain with the same orientation as the triangular
faces of the tetrahedra at the ends of the virtual link. Thus, the orientation
propagates through the chain.
In figure 2.9 we summarize which elements can be connected using the above
algorithms. A solid line between elements implies we can connect a given mesh made
of these two element types with no problems. If the line is dashed then the mesh may
have to be changed to meet the connectivity constraints.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
28
£
ir
Figure 2.9: Connectivity summary. Solid arrows imply we can connect elements together with no extra mesh refinement, dashed arrows mean that there are some constraints on allowable configurations.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter 3
B asis Functions
This chapter is devoted to outlining the polynomial bases tha t we use to represent a
function defined on an element. A desirable basis has three properties: (i) it should
be orthogonal or nearly orthogonal under a convenient norm, (ii) well behaved a t
high polynomial order, and (iii) it should be computationally efficient to take inner
products w ith this basis if it is to be used in a Galerkin framework. The bases we
present are designed to satisfy these conditions. The computational complexity of
taking inner products has been minimized for these hierarchical bases by insisting
that each basis function is a tensor product of one-dimensional shape functions.
3.1 One-Dimensional Orthogonal Polynomials
In the following sections we use make considerable use of the Jacobi polynomials.
These are defined as the polynomial solutions of the following Sturm-Liouville prob-
29
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
30lem:
(1 - * ) ( ! + x ) ^ ^ - + p - a) - (2 + a + /3)x\ 5y(x)dx dx= -A ny(x)
or
d_dx
(1 -a : )1+Q( H - 2r)1+ dy(x)dx = —A„(l — x )a{l + x)py(x)
An = n(n + a + /? + 1)
y(«) =
They have the following orthogonality:
/ ( 1 — rr)Q(l + x )^ P ^ ,0( x ) P ^ ( x ) d x = 0 n ^ m
J \ 1 - x)a(l + x)<3P ^ { x ) P ^ { x ) d x =
qcc+p+i r(n + a + l)r(n + 0 + 1)2n + a + ft + 1 n !r(n -+- a + j3 + 1)
As the following sections were just being completed I found th a t work along exactly
the same lines had just been subm itted to the SIAM Journal on Numerical Analysis
by W ingate and Taylor [28]. I include the work here for completeness noting that the
results have already been independently obtained and presented there, except for the
results based on prisms.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
313.2 Singular Sturm-Liouville Problems on a Sim
plex in d-Dimensions
The following section was inspired by the work of Owens [29]. In th a t paper he
derived an orthogonal basis for approximating functions on a triangle from a basic
premise. We will take an alternative route and show that it is possible to find a
singular Sturm-Liouville operator for a d-dimensional simplex and present its eigen
function/eigenvalue pairs explicitly. The eigenfunctions for the cases d = l (a segment),
d=2 (a triangle), and d=3 (a tetrahedron) will prove to be useful in the context of
polynomial-based approximation on these simplices. As stated before these results
have been proposed in [28].
A d-dimensional simplex S d can be defined as a set of constraints on the entries
of a d-dimensional vector:
S d = {r 6 Rd\0 < < 1; j = 1 ,2...,d}i=1
We define an operator on a d-dimensional space of at least twice differentiable func
tions of d variables:
L t = J 2 dn (rA - ~ ri S rA J )i=i j =i
We also define a (d + 1) vector, s = [(1 — u))ri, (1 — uj)t2 , (1 — w )^ , n>], where
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
u) € [—1,1]. Using the identities:
dsi — - dr. i = l , d1 — CJ
1 i=ddsd+i = Y v j d ri 4- dy
L ~ u i=i
Sid Si = Tidri i = 1, di=d+1 i i=dY S*dsi = i----- Y Tidn + Udwi=l 1 ~ W i=l
it is straightforward to show the following relationship:
From this we can repeat the recurrence relation ending up with the operator in terms
of the a coordinates:
L* = £ rTj=d r [5«i(oi(l - ai)d<n) ~ ~ 1) 0,-]1=1 Uj'=i+l(l
where a is defined by the canonical transform:
r i = a i( l - a2)( l - a 3)...( l - ad)
r2 = a2( l - a 3) . . . ( l - a d)
rd — ad
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
33This form of the operator shows tha t L \f is self-adjoint in the inner product taken
over the simplex S d. This is because S d maps to a d-dimensional unit box Ud = {r G
Rd|0 < Ti < 1 i = 1 , 2 and using integration by parts we notice that all the
surface integral terms are zero.
For example, we consider the ith operator in the sum for L d:
ud
ud
(u,(LiYv)st = (u, n jg i1 - O . W - (i -
, — a k ) k~ l f a I M ^= (W> ( t t jW ------- r ~ ) L { “ a*') v h do-j)
= (^ (O ifl - a,.)%)] u, at).<:~1 )v)u-
= ((^r Yu,v)s<l
We can also use the definition of the Jacobi polynomials to show that Ld has
eigenfunctions:
</>i = H S f ( l - Gi)*Pd?*'~L'°(Oi)
where d. € Nd, a = dj. The eigenvalues for these eigenfunctions are:
= cd+i (cd+i + d)
In summary, the operators we have defined in the d-dimensional simplex have simple
tensor product eigenfunctions when mapped to a d-dimensional unit box. Also, their
eigenvalues are n(n + d) where n is the total degree of the eigenfunction in both the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
34mapped and original coordinates a and r.
3.3 Orthogonal (Modal) Bases
The results we now present for the triangle and the tetrahedron are special cases of
the previous section, but for clarity we will outline some of the steps used to find
these results.
We first consider the tetrahedron as this has the most complex mapping between
the local cartesian coordinates and the tensor coordinates. Using the scaled coordi
nates (a, b, c) E [0,1] x [0,1] x [0,1] for the tetrahedron (for simplicity) we can express
the local orthogonal coordinates (r, s, t) E {0 < r, s, t, r + s -f- £ < 1} as:
r = a ( l — 6)(1 — c)
s — 6(1 — c)
t = c
We consider the operator:
Lret ■= dr[r{ 1 — r)dr - rsds - rtdt) +
ds(s (l — s)ds — srdr — stdt) +
dt(t( l — t)dt — trdr — tsd3)
Using the following identities it is straightforward to express the operator with respect
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
35to the (a, b, c) coordinates:
rdr = ad,a
sd3 = ~ ab da + bdb (1 - b)
idt = 7= jtt:------r<9a + ■>- ~- - vdb + ed c( l - 6 ) ( l - c ) (1 - c )
rdr 4- sds + tdt = T 7771 + 77— r d 6 + cdc( l - 6) ( l - c ) (1 - c )
After some m anipulation the operator can be expressed in terms of the (a, 6, c) coor
dinates and is:
= (1 - i) ( 1 1 c) [ d M l - + ( ! --->) (1 - c) [at(6(1 - 6)2ft)]
+ ( 1 3 ^ 2 f c M 1 - c)3 ) ]
This demonstrates tha t the operator maintains tensor form in the (a, b, c) coordinate
system. It is now trivial to show that this is a self-adjoint operator by applying one
dimensional integration by parts to each of the three tensor parts. Also, by using the
definition of the Jacobi polynomials we can show that the orthogonal basis is a set of
eigenfunctions of L r et and find their eigenvalues.
We will now show th a t the polynomial functions 4>ijk defined by:
4>ijk = P , '0 (a) (—t — (b) (—r — P^ t+J'+2,0 (c)
are eigenfunctions of the Lret operator. We consider the first part of the operator.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
36The definition of the Jacobi polynomial directly implies the following relationship:
da{a{ 1 - a)da(f>ijk) = -* (t + 1 )& #
We now consider the first two terms of the operator. Using the previous result we can
remove the dependency on a and then use the definition of the Jacobi polynomials
with non-zero (a, (3) to show that the polynomials are indeed eigenfunctions of the
first two terms of the operator:
Lbfiijk — bdb(f>ijk
i(i + 1)
1 - 6O.0/ ^ D2(*+i)+2,0= ( i - b y p r ( a ) P { (c)
b ( l — + (^ — (3 + + (—*(2 -1- z))
= 4>ijk [~ j ( j + (2z + 1) 4- 1) — z’(2 + *')]
Pi2i+l,0
(6)
Applying the same technique again we come to the relationship:
Lret&jk = dc(c( 1 - c)dc<f>ijk) ~ 2cdc<t>ijk - + + 2l(pijk1 — c
= Kjk4>ijk
Thus, the tensor product of Jacobi polynomials <j>ijk are eigenfunctions of the total
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
37operator L ret with eigenvalues:
Aijk = —{i + j + k)(i + j + k 3)
This approach clearly applies to the triangular, quadrilateral, prismatic and hexahe
dron elements as well.
3.3 .1 T he O rth ogon al Triangle B asis
The triangle is a special case of the tetrahedron with c = — 1. As before, we can
specify a self-adjoint operator tha t has the orthogonal basis, proposed by Proriol [17]
and later by Koomwinder [18] then Dubiner [16], as its eigenfunctions. We set out
the scaled coordinate system, operator and eigenfunction/eigenvalue pairs:
r = a(l — b)
s = b
Lth = dr (r( 1 - r)dr - rsds) + ds(s( 1 - s)d s ) - rsd r )
= r ^ [a“ (a(1_a)a<-)1 + ( r = T ) 56(i’( 1 - i ’)2s‘)
= Pi'Q(a)(l — b)'Pji+lfi(b)
A ij = — (i + j ) (i + j + 2)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
383.3.2 T h e O rthogonal Q uadrilateral B asis
The straight tensor product property of the quadrilateral gives a very simple form
for the operator, it is the sum of two one-dimensional operators. Similarly, each
eigenfunction is a straight tensor product of one-dimensional eigenfunctions and their
eigenvalue is a sum of the one-dimensional eigenvaiues.
r = a
s = b
LQuad = dr{r{ 1 - r)dr) + ds (s( 1 - s)3,))
= da{a{ 1 - a)da) + db(b( 1 — b)db)
toj = Pi'°(a)Pj'a(b)
Xij = —i(i + 1) - j ( j + 1)
3.3.3 T h e O rthogonal P rism B asis
The prism is simply a tensor product of a triangle and a uniform third direction.
Hence, we sum the operator for the triangle in th e (r, t) directions and the one
dimensional operator in the s direction. Likewise, we can obtain the eigenfunc
tion/eigenvalue pairs from the above analysis for th e triangles.
r = a ( l — c)
s = b
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
39t = c
Lprism = dr(r( l — r)dr - rtdt) + d3(s ( l - s)ds) + dt {t(l — t)dt — rtdr)
= 3^ [8 «(o(l - a)a„)] + [%(6 ( 1 - 6 )%)] + 3^ [ScWl - o)2Sc)]
<hk = P i - ° ( a ) P ^ ( b ) ( l - c Y P ^ +l’a(c)
Xijk = — (i + k) (z + k + 2) — j ( j -I- 1)
3.3.4 T h e O rthogonal H exah ed ral B asis
The hexahedral analysis is trivial since, like the quadrilateral, the hexahedron is
a straight tensor product of three one-dimensional directions. The operators and
eigenfunction and eigenvalue pairs are:
r = a
s — b
t = c
Ttfex = dr(r( l - r)dT) + ds(s ( l - s)ds) + dt(t(l - t)dt)
= da(a( 1 - a)da) + db(b( 1 — b)db) + <9c(c(l — c)dc)
<f>ijk = P °’° (a) Pj'° (b) P °’° (c)
Xijk = + 1) - j ( j + 1) - k(k + 1)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
403.3.5 T h e D eca y o f B asis C oefficien ts
Thus, each basis is a set of eigenfunctions of a singular Sturm-Liouville operator which
leads us to the following observations:
&ijk =" (u, 4>ijk)
= (w> -r— Hfajk))■ijk
''■ijk
ijk
Hence, if the function is infinitely smooth we see tha t the coefficients Uijk must de
crease faster than any polynomial power of i , j , k. Thus, the sum:
N N N
«* = E E E « ijkfajki j k
must converge exponentially fast to u as N increases for all infinitely smooth u.
It is im portant to notice that since straight-sided tetrahedra and triangles have
constant geometric mapping Jacobians these results hold for arbitrarily stretched
tetrahedra and triangles. This does not follow for the other elements since their
geometric Jacobians are quadratic for non-perpendicular elements. This backs up the
findings th a t the simplical elements handle deformation better than the other types.
Unfortunately, it does not appear th a t this method generalizes to the pyramid in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
41a straightforward way. However, a suitable orthogonal basis is known for a pyramid:
r = a (l — c)
s = 6(1 — c)
t = c
frjk = Pi'°{a)Pj'°{b)(l - c)i+jP 2k{i+j]+2'\c )
This basis is only appropriate for supporting P n since the ‘c’ component is of the
order i + j + k. Thus, if each i, j , k < N but i + j + k > N then it is necessary to use
high-order quadrature to integrate these modes exactly. Also, if i + j + k < N the
function is a polynomial in r, s and t.
3.4 Semi-Orthogonal Bases
In appendix A there is a summary of the modal triangle basis proposed by Dubiner
[16] and the tetrahedral basis proposed and investigated by Sherwin [30]. Also, we
list the modal bases for the prism, pyramid and hexahedron suggested by Sherwin in
[27] and with Karniadakis in [31]. In the following subsections we will examine the
properties of the bases for the triangle and quadrilateral. We wall also propose some
modifications to the shape functions that will enhance efficiency for the quadrilateral
and provide compatibility between a triangle and a nodal quadrilateral.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
423.4.1 M od a l basis
A modal basis for triangle elements has been presented in [16] and [21], and it was
applied to to fluid dynamics problems in [20], and to geophysical fluid dynamics
problems in [32]. We now consider a basis for quadrilaterals which is compatible with
the triangle basis and thus they can be used together. This combination was first
proposed in [27] for a set of three-dimensional polyhedra.
T riang le B asis: T rH H
In this and the following sections we will define a set of bases for the triangle and the
quadrilateral. For clarity we will label them with the first two letters of the element
they will be used on (i.e. Tr or Qu for triangle and quadrilateral respectively). The
next letter will denote the type of mode shapes used on the boundary of the element.
H will denote hierarchical and N will denote non-hierarchical. The last letter will
denote the type of mode shapes used for the interior modes, again using N and H as
before. For example, T rN H will denote a triangle T r with non-hierarchical boundary
modes N and hierarchical interior modes H.
We present here a basis which is a set of tensor products with respect to the tensor
product coordinates for the triangle and polynomials with respect to the reference
element coordinates. It maintains numerical linear independence up to high orders
due to the construction of the interior modes from Jacobi polynomials with carefully
chosen (a, /3) coefficients to ensure that mode shapes do not become too similar.
Increasing a shifts the roots of the Jacobi polynomials away from the coordinate
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
43singularity a t 6 = 1 as dem onstrated in [21] and hence the modes are prevented from
having the same shape a t this vertex.
The form of the T rH H basis is:
Vertex Modes:
T — a \ ( l — biv e r te x i _ f ^ V 2 J \ 2
(pvertex2 ^ ^
<f>vertex 3
2 J \ 2 '1 + 6'
Edge Modes (2 < m; 1 < n, m < M\ m + n < N ):
1 - b \ m= ( ^ ( ^ ) ^ ( ° > ( 2 ;
= ( ^ ) ( ^ ) ( ^ ) *£!,(»)
«*- - ( ^ ) ( ^ ( ^ ) p"->>
Interior Modes (2 < m; 1 < n, m < M] m + n < N):
= (^ ) (^ ) («) ( ^ ) ”“ (^ ) (»)
We represent this basis graphically for N = 5 in figure 3.1; the highest mode is
quartic.
In the next chapter we will be considering the variational form of elliptic operators.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
44
\.o050.0
Vertex ModeInterior M<
Figure 3.1: Mode shapes for the triangle modal basis (T rH H ) with N = 5.
At this point it is useful to note tha t the T rH H basis has a very useful property that
the bandwidth of its stiffness m atrix is N + l if the modes are ordered in a specific
way. The stiffness m atrix is defined as the inner product of the the gradients of the
modes i.e.
where the 0t- represent the modes indexed in a given way. This indexing is important.
If we list the vertex modes, then the edge modes and then the interior modes we see
in figure 3.2 that the stiffness m atrix is partially banded. The interior modes in this
case have bandwidth 2N + 1 as shown in [30]. However, if we list the modes lowest
order first, i.e. linear modes, then quadratic modes, then cubic and so on, we see in
figure 3.3 tha t the stiffness m atrix becomes strongly banded with bandwidth N + 1.
Sij = (V0t-, V0j)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
45This will become im portant later on when we need to invert matrices of this type
since reducing the bandwidth reduces the storage and number of operations required.
Q u a d rila te ra l B ases: Q u H H an d Q u H N
For quadrilateral elements Q l is a bijection so we do not need to worry about the
coordinate singularity as in the triangle case. We are free to choose a set of modes that
are C° compatible with the triangle expansion. An obvious choice th a t guarantees a
high degree of orthogonality is the Q u H H basis:
Vertex Modes:
Edge Modes (2 < n , m < N):
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
46
Figure 3.2: Top: Geometric ordering for modes, Bottom: Stiffness m atrix (V<fo, Vdj) for the continuous basis for triangles (iV = 15) with the geometric ordering of the rows and columns.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
47
! I W { S|§lfe|fei>
Vertex Modes □* Edge 1 Modes
= Edge 2 Modes
= Edge 3 Modes
□ : Interior Modes
10
2030
40
50
60
70
8090
100
11 0
50 1000
Figure 3.3: Top: Polynomial degree ordering for modes, Bottom: Stiffness matrix (V0,-, V4>j) for the continuous basis for triangles ( N = 15) with modes ordered by polynomial order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
48Interior Modes (2 < m, n < N):
in teriorm n
We represent this basis for N = 5 in figure 3.4.
.Vertex Mode
Figure 3.4: Mode shapes for the quadrilateral modal basis (Q uH H ) with N = 5.
Alternatively, we could also choose a second set of modes with an even better
orthogonality relationship. To this end, we use the Legendre interpolant functions,
which we will investigate more thoroughly in the next section. These nodes are used
in a collocation manner to maximize the discrete orthogonality of the modes in the
interior region. More specifically, the Legendre quadrature points are used as nodal
points as well. We also modify the vertex modes and replace the edge modes with
tensor products of the one-dimensional modal basis and the Legendre basis. Similarly,
we replace the interior modes with tensor products of only the Legendre basis. This
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
49means th a t edge modes from one edge are orthogonal to edge modes from another edge
and all the interior modes. The interior modes are m utually orthogonal and hence we
can perform inner products and backwards transforms in 0 ( N 2) operations. Before
we write the basis we need to define an appropriate Lagrange interpolant in terms of
the Legendre polynomial Pn(x) = P%'° as:
M r) = d - ’- W ON ( N + l)Pjv(rn)(r — rn)
where r t- denote location of roots of (1 — r2)P'N(r) = 0 in the interval [—1,1]. Also,
by definition we have th a t hn(rm) = Smn. This construction destroys the hierarchy
of the basis in the interior of the element. More specifically, this Q uH N basis is
defined as follows:
Vertex Modes:
< r rteI = hda) +ft,(6) ( i ^ ) - hi (a) hi (b)
<Tr‘“ s = h„(a) - hN(a)hi(b)
= hH(a) + M &) ( V p ) - h „ ( a j h N(b)
= ft,(a) + ft„(f,) (L Z _?) - h M h s i b )
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
50Edge Modes (2 < n, m < N):
V S T = ( ^ p )
V*™ = hN(a) pM ^J)
« s=s = (^p) ( ^ ) P ^- i (a )h N (b)
t i d" ' = M&) ( ^ ) ( ^ p ) p» i W
Interior Modes (2 < m , n < N):
= hm(a)hn(b)
i.o
os 0.0
Figure 3.5: Mode shapes for the quadrilateral mixed modal basis Q u H N with N = 5.
Note tha t according to our convention we have th a t hi(—1) = 1 and h ^ ( l ) = 1,
etc. In figure 3.5 we see that the edge modes now have behavior localized to the edge,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Vertex M ode
51and the same is true for the vertex modes. In figure 3.6 we present graphically the
inner product (0mn, <ppq) which is the mass m atrix for this basis. Here, we order the
vertices first (around the origin), followed by the edges, and the interior contributions.
We verify tha t indeed there exists great sparsity in this m atrix indicative of the strong
discrete orthogonality between modes.
100
150
200
0 200100
Figure 3.6: Mass m atrix for the quadrilateral mixed modal basis (Q uH N ) with N = 15.
3.4 .2 N o d a l and M ixed B ases
We have so far presented a hybrid basis derived from Dubiner’s [16] C° basis for the
triangle. This dictates the shapes of the edge and vertex modes for the quadrilateral.
We will now consider an alternative starting point. In the spectral element method
[7], quadrilateral elements are commonly used with a nodal basis, constructed from
the Legendre interpolant polynomials. This method benefits from a fast transform to
the polynomial space, due to the collocation property of the modes. Similarly, inner
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
52products can be evaluated efficiently.
Q uadrilateral Basis: Q uN N
This basis is constructed based on the Gauss-Lobatto-Legendre interpolant of order
N already defined as {hn(r)}. It was first used for spectral element discretizations in
[33]. This is a nodal non-hierarchical basis.
For consistency we present the QuN Nbasis in the same format as the hierarchical
basis:
Vertex Modes:
^ v e r te x , =
= h ^ h ^ b )
0 vertex3 = ^ ( a) ^ ( 6)
^ v e r te x , = h ^ a ) h [ f ( h )
Edge Modes (1 < n ,m < N):
^ = h M h ^ b )
t f 9*2 = hN(a)hn(b)
t f e3 =
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
= hi(a)hn(b)53
Interior Modes (1 < n, m < N):
C r ,0r = hmWhnib)
We represent this basis for N = 5 in figure 3.7; all modes are fourth degree polyno
mials.
,Vertex Mode
Figure 3.7: Mode shapes for the quadrilateral nodal basis (Q uN N ) with N = 5.
T rian g le B asis : T rN H
We will discuss four ways that the triangle basis can be constructed, but eliminate
three of them due to undesirable properties of the resulting basis. A first approach
might be to construct the triangle basis in the same way as the quadrilateral, using a
complete set of N 2 tensor products of the Legendre interpolant functions {/in}. This
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
54basis suffers from over-resolution a t the singular vertex of the coordinate system as
shown in [16]. To avoid this overresolution we could choose a subset of the Legendre
interpolant functions on T2, as suggested in [34]. This leads however to a basis that
suffers from numerical linear dependency as noted in [16]. Lastly, we could construct
a two-dimensional nodal basis in the manner of [35] or [36] by explicitly specifying
the node distribution, but these bases do not have the tensor product property thus
making spatial derivatives expensive 0 ( N 4) operations. Recently, however, fast polar
derivative techniques have been developed [37] that reduce the scaling factor in the
asymptotic cost. These advances have made the cost comparable to the tensor product
cost for N < 10. It is still true tha t a modal approach provides a natural way to to
use exact integration and thus avoid aliasing.
Instead, we propose a basis th a t is compatible w ith the nodal quadrilaterals
(Q uN N ) and numerically similar to the modal triangle basis (T rH H ). The C°
continuity condition requires that the boundary modes of the triangle must have the
same shape as the {hn} modes of the quadrilaterals they can share an edge with.
However, this condition does not determine how the modes should be shaped in the
interior of the triangle, which leaves us a certain amount of freedom. We require
that the modes axe polynomials in (r, s) and that the basis spans the polynomial V m-
We complete the requirement th a t the bases be numerically similar by ensuring the
new basis is numerically linearly independent. This is guaranteed by using the same
interior modes as the modal triangle basis. As previously discussed, the choice of
(a , /?) for the Jacobi polynomials is key in ensuring th a t the interior modes do not
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
55
degenerate in the (r, s) coordinate system but are still tensor products in the (a, b)
coordinates.
We constructed the vertex and edge modes of the new basis in the (r, s) coordi
nates. The vertex 3, edge 2, edge 3, and interior modes still m aintain tensor-product
in the (a, b) coordinates but the remaining boundary modes become two-dimensional
functions in this frame. These new modes will increase the constant factor in the
asymptotic cost of inner products and transforms compared to the modal triangle.
The new basis is presented in closed form, but we note th a t it could also be derived
as a linear combination of modes from the modal basis, and this is how we guarantee
th a t it will share all properties that the modal basis has for linear operations.
So we now present the nodal-compatible T rN H basis for the triangle in local
cartesian coordinates: Vertex Modes:
^vertex i = ^ ^ + g + ^
^vertex* = hff ^
0 vertex3 = ^
Edge Modes (2 < m, n < N):
/ r + s') U - r j
) hm_ i( r )
(j dge 2 _ f l +U - s y ) hn—i (s)
C *363 = - i (S ) 1 ^n—1 (^)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
In standard triangular coordinates the vertex Modes are:56
^ ertexi =
(a ( 1 ^ ) - (1 ± * ) )
4>verteX3 = htf(b)
Edge Modes (2 < m , n < N):
= (1 - q )(l - ft) , ( J l ^ ± \ _ (4 — (1 + a) (1 — 6) m_1(, 2 J 2 ,
^ = ( —J —) ^>.-1 M
The interior modes are unchanged from the modal triangle basis. We represent this
basis for IV = 5 in figure 3.8; the interior modes still have the bubble shape and thus
they are zero at the boundaries.
3.5 Comparison of the Bases
We have presented two classes of bases in sections 3.4.1 and 3.4.2 that give the same
approximations and share the same spectral properties. We are left with the decision
of which basis should we use for a given problem. In this section we summarize their
properties and suggest possible selection criteria.
First, we consider the structure of the mass and stiffness matrices encountered in
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
57
Vertex Mode
Vertex *3*
Vertex *2*
Vertex *1
Figure 3.8: Mode shapes for the triangle mixed basis (T rN H ) with N = 5.
the convection and diffusion equations, respectively. We concentrate on the triangular
elements only, and we examine the different structures corresponding to bases T rH H
and T rN H . The elemental mass m atrix is the m atrix which has entries defined by:
Bmn — (jfimi < n) •
Figure 3.9 shows in (a) the mass m atrix for the T rH H basis, and in (b) the mass
m atrix for T rN H and in (c) the mass m atrix corresponding to T rN H but calculated
using exact integration. In the latter case the number of quadrature points needed
is larger than the nodal points unlike the case in (b) where the quadrature points
coincide with the nodes a t the edges. We notice that the sparsity of the mass matrices
corresponding to bases represented in (b) and (c) is reduced compared to the T rH H
modal basis, but this is to be expected because each boundary mode in T rN H is a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
58linear combination of all of the T rH H boundary modes.
(a)
(c)
Figure 3.9: Comparison of the elemental mass m atrix for (a) T rH H , (b) T rN H using collocation, and (c) T rN H using exact integration for N = 15.
We can also compare the elemental stiffness matrix for the TrH H m odal basis and
the T rN H mixed basis. The stiffness matrix is the m atrix which has entries defined
by:
Smn = (V0771) .
Figure 3.10 shows the two stiffness matrices for N = 15. Again the matrix corre
sponding to T rN H basis has a denser structure.
We now summarize the properties of the two classes of bases. The modal basis
(T rH H and Q uH H ) properties are:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
59
(a)
: w xsss
Figure 3.10: Compaxison of the elemental stiffness matrix for (a) TrHH, and (b) TrNH bases for N = 15.
• The basis is hierarchical.
• Gaussian integration order is not dictated by the basis order.
• We are able to vary locally the number of modes per edge or interior.
• Elemental transforms and inner products are 0 ( N 3) operations.
• The interior-interior mass matrix is banded.
• The interior-interior stiffness matrix is banded.
In addition, this basis can be enhanced with other properties by varying the form of
Jacobi polynomials. For example, an interesting version proposed in [16], [32] uses
(a, @) = (2,2) and (2 m + 3 ,3) for the Jacobi constants. It has the following properties:
• The interior-interior mass matrix is diagonal.
• The interior-interior stiffness matrix is full.
There are many different choices in choosing a mixed basis. Their main properties
are:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
60• The basis is non-hierarchical (Q uH N ,TrN H , and Q uN N ).
• Gaussian integration order is dictated by the basis order (Q uN N , TrN H a t
the edges, and Q uH N in the interior).
• The expansion order is (practically) fixed in nodal quadrilaterals (Q uN N ).
• Quadrilateral transforms and inner products cost 0 ( N 2) operations for Q u N N
and 0 ( m 2) for Q uH N .
• Triangle transforms and inner products are 0 ( 2 N 3) operations (TrNH).
• The interior-interior mass matrix is banded (TrNH); it is extremely sparse for
Q uH N and diagonal for QuNN.
• The interior-interior stiffness matrix is banded only for T rN H , otherwise it is
full.
These properties suggest situations where each type of basis is appropriate: For
example, the m odal basis can lead to high com putational efficiencies, if:
• The solution has local regions of interesting behaviour.
• The solution benefits from non-steady regions of interesting behaviour - this
can be captured by local p-refinement.
• The domain is highly irregular and needs a high ratio of triangles to quadrilat
erals.
On the other hand, the mixed basis can also lead to high efficiencies, if:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
61• The solution, has a uniform variability.
• The domain is locally irregular, and thus it does not require a large number of
triangles to complement the quadrilaterals.
More importantly, it is the specific application th a t we consider and the dynamic
refinement procedure th a t ultimately decides which basis functions are the best choice.
If it becomes apparent tha t an initial choice of basis becomes inappropriate, it is not
an expensive operation to transform between bases.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 4
O perators
Now that we have set up the basic geometric properties of the elements, and how
to represent functions on them the next step is to show how their data can be ma
nipulated economically. In this chapter we develop a method for efficiently imple
menting the family of hybrid elements on a computer. Specifically we present an
efficient method for evaluating surface integrals on the elements. Also we examine
the approximation and wave propagation properties of these elements experimentally.
Finally, we provide a simple algorithm for minimizing the bandwidth of the matrices
that result from the construction of discrete linear operators for a domain.
4.1 Elemental Spatial Operators
In this section we consider the two basic operations that will be performed on a
function represented on an element. These are taking the spatial derivatives and
integrals of functions represented on each element type.
62
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
634.1 .1 In tegration
We take advantage of the tensor product element coordinate systems to perform
integration. The integrations over each element can be performed as a set of one
dimensional integrals using Gauss quadrature. If we used the reference coordinate
systems this would be very difficult since the limits of the “collapsed” elements are
not constant.
We first describe the choice of quadrature type for integrating each direction. We
will then motivate the inclusion of quadrature with non-constant weights in order to
reduce the number of points we use.
In two-dimensions we consider integrals of the form :
f f ( x )d x d y = f / ( x ( r ) ) ^ y c £ r d sJ P hysical J R eference u ( T J
= f /(x (r (a ) ) )9(-*> i f ^ dadbJ T e n s o r ^ ( T /
In three-dimensions:
/L f (* )d xd yd z = f / ( x ( r ) ) ^ ^J P h y s ic a l J R e fe r e n c e
drdsdt
= [ /(x (r(a )))JT ensor
d(x) d(v) dadbdcJTensor
We use the Gauss weights th a t will perform the discrete integral of a function as a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
64
Element ‘a ’ ‘b ’ ‘c’Triangle G LL GRJifi -
Quadrilateral G LL GLL -Tetrahedron G LL GRJi'O GRJ2to
Pyramid G LL GLL GRJ2,oPrism G LL GLL G R J1jo
Hexahedron G LL GLL GLL
Table 4.1: G LL implies Gauss Lobatto Legendre which is the Gauss quadrature for a constant weight function with both x = ± 1 points endpoints included. G RJQip implies Gauss Radau Jacobi quadrature with (a, (3) weights and the endpoint x = — 1 included.
sum:
(1 - z)"(l + z)Bf ( z ) d z = Y ,J~l i=0
This will be used in each of the d directions in the d-dimensional elements. In table
4.1 we show the type of Gaussian quadrature we use in each of the ‘a ’,‘b ’ and ‘c’
directions.
4.1.2 F lu x Integrals
In the following chapters we will outline the Discontinuous Galerkin Method. In this
method it is necessary to evaluate terms of the form:
where dQ is the boundary of an element f2, for all the (f>n test functions in the elemental
basis. There are test functions for a triangle so the boundary integral is an
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
650 ( N 3) operation. This means that the flux integration is as expensive as the volume
integral. We can reduce the cost of this integral by examining the discrete sum form:
f M = Y^4>n(ai,0)fl (ai)wfJl (ai )+'^2(f)n( l ,bi ) f2(bi)WiJ2(bi)J a n i= 0 i= 0
t = 0
where J n and f n are the Jacobian and flux function for the nth edge.
We can rewrite the edgei flux as:
A , /&> = EJedgei j =Q i=Q W 0
where:
0 if j5{j —
1 if i = j
The fluxes for the other edges can be constructed in similar ways. Using this summa
tion representation we can now evaluate the surface flux integral by adding the edge
fluxes scaled by weight and Jacobians to the F field and then evaluating one volume
integral.
4.1 .3 D ifferen tia tion
A function represented at the quadrature points can be considered as a sum: if /,-j is
the (i , j ) th physical value of the function then the interpolating polynomial is:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
66
f M = ' t ' t f ijht(a)hbj(b)i j
Here /if is the ith interpolating polynomial for the quadrature chosen in the ‘a ’ direc
tion. Thus, we can evaluate the derivative of this interpolating polynomial in each of
the (a, b) directions.
daf(a ,b ) = E E W M ) ( a)/lK i )n m
We can now evaluate the (i , j ) th quadrature value of the derivative as a m atrix vector
product:
daRai, bj) = J 2 f n i da(hn)(ai)n
Repeating the above we can evaluate the ‘b ’ derivative (and the ‘c’ derivative in
three-dimensions). To calculate the x ,y ov z derivatives we simply apply the chain
rule. In implementation we store the geometric factors for all the elements. This is
more memory efficient for straight-sided triangles and tetrahedra since the geometric
factors are a set of constants (except for (1 — b) and (1 — 6)(1 — c)2 factors).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
674 .1 .4 Im p lem en tation and P rogram m in g
So fax we have outlined specifications for each element type. Each description includes
three coordinate systems, a set of basis functions, a set of operators and a set of surface
panels. By carefully defining each element type this way we can implement each type
using a unified methodology. We used an object oriented programming approach.
The first task is to define a “generic” element that has a sufficiently flexible da ta
structure. The computer language C+-F allowed us to define a base Element class.
The class was given storage for the following:
• The geometry of the element (i.e. the vertex positions and curvature informa
tion).
• The connections the element makes to other elements, through vertices, edges
and faces.
• Storage for the values of the field at the element’s quadrature points.
• Storage for the basis shape function coefficients.
• Parameters determining the expansion order for edges, faces and interior modes.
• A look up table for the functions (operators) defined on the element and its
data.
The last item is im portant for efficiency. If we were to create a class th a t did not
know how to operate on its own data, then we would need a list of conditional “i f
statements every time such an operation needed to be done in the code. Clearly this is
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
68not an efficient approach. The object oriented programming (OOP) approach replaces
the set of i f statements with a memory access that determines which operator should
be used for the element in question. Also this method is efficient for progra m m in g
purposes. Once an operator is defined for all the elements, any piece of code that
requires this operator only has to call the one interface to the operator regardless of
the element type. For instance, say we would like to differentiate a function held in
an array of values at quadrature points {field) then in pseudo code:
r y-r j- j.t t- u d f ie ld d f ie ld d fie ld .element— > Gradienti field, — ---- , —----- , — -----)ox ay oz
would take the spatial derivatives of field and place the x, y and z derivatives in the
arrays named S i m , and
A more complicated example might be to create two elements, set the data values
at the quadrature points of the elements, differentiate the fields and compare the
differentiation errors on each element. The pseudo code to set up the elements is
shown in figure 4.1. We would like to point out th a t the implementation of the
element constructors is provided in the library routines of A/’s k 'T'ocr . In figure 4.2
we show the code for the next stage of applying the derivative operators to the list
of elements we have created. Notice in this section th a t no explicit mention of the
types of each element is made when we use the operator on them.
In the next section we consider the discretization of the projection, the convection,
and the diffusion linear operators. They are formulated as discrete operators in each
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
69
/ / Make an array to hold the elements Element *elements[2];
/ / Set up the coordinates for a triangle Coordinate tri-coordinates; tri_coordinates.x = {-1, 1,-1}; tri_co ordinates.y = {-1,-1, 1};
/ / Set up the degree vector for the triangle int tri_degree_vector[4] = {2,2,2,!};
/ / Create the triangleTri elements[0] = new Tri(tri_degree_vector, tri_coordinates);
/ / Set up the coordinates for a quadrilateral Coordinate quad_coordinates; quad_coordinates.x = {-1, 1, 1, -1}; quad_coordinates.y = {-1,-1, 1, 1};
/ / Set up the degree vector for the quadrilateral int quad_degree_vector[5] = {2,2,2,2,1};
/ / Create the quadrilateralQuad elements[l] = new Quad(quad_degree_vector, quad_coordinates);
Figure 4.1: A piece of M e n T a r code demonstrating how a list of elements can be constructed.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
70
/ / Set the quadrature point values of both elements for( i = 0; i < 2; i=i-Fl )
elements [i] — > Set-field ( “sin (PI*x) *cos (PI*y) ” );
/ / Create some storage for the derivatives of the fieldsdouble* data[2];for( i = 0; i < 2; i= i+ l ){
dx[i] = elements[i]— >Duplicate_data-space(); dy[i] = elements[i]— >Duplicate_data_space();
}/ / Take the spatial derivatives of the fields we just set for( i = 0; i < 2; i=i-f-l )
elementsjj]— >Gradient(elements[i]— >data, dx[i], dy[i])
/ / Calculate the error norm of the ‘x’ derivatives and the exact answer for( i = 0; i < 2; i= i+ l )
elements[i]— >Error(dx[i], “PI*cos(PI*x)*cos(PI*y)”);
Figure 4.2: A piece of A fe tt 'T c tr code demonstrating how operations can be made on a list of elements.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
element and then globally assembled over the entire domain.71
4.2 Global Projection Operator
We construct the Galerkin projection that minimizes the L2 error for approximation
of a function by the C° basis. F irst we define the elemental Jacobi transform from
the polynomial space to the physical space:
N
U ~ ^ 'U-n<Pn
where un is the coefficient of the n th basis function <f>n. From this we can obtain an
approximation for the coefficients un. If we take the inner product of both side with
<j)m we obtain:
N
(0 rrn w) = ^ I ^ m i
which we can solve for un since (<£m, 4>n) is positive definite [20]. Explicitly, we can
write the projection coefficients as:
= B mn (0m w)
where B is the m atrix with entries B mn = (<£m, 4>n)- This approximation minimizes
the residual:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
72
N
r — 11 y i UTnfimWL? ■
4.2.1 G lob a l C onvergence in Skew E lem en ts
A: D:
G: H:
Figure 4.3: Meshes (A-H) are consisted of three quadrilaterals and two triangles which are progressively skewed by shifting the interior vertex.
ofcW
10* 10*
10*
10* 10*
10* 10* 10* r
10'
10"10"10"10"
11 13
Expansion Order
Mesh A Mesh B MeshC Mesh D MeshE Mesh F Mesh G Mesh H
15 17 19
Figure 4.4: Convergence in the L2 norm for modal projection of the function u sin('Kx)sin(jvy) on meshes A — H.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
73
10"
10"
10"
10"
10"
o 10'Tf r l 1 0 "
1 0 ™
1 0 "
1 0 ' “ r
10"'10’ ’
Mesh A Mesh B MeshC Mesh D Mesh E Mesh F MeshG Mesh H
7 9 11 13
Expansion Order
-1—015 17 19
Figure 4.5: Convergence in the L 2 norm for mixed, projection of the function u = sin(irx)sin(Try) on meshes A — H.
We now consider the effect of the skewness of the physical elements on the accuracy
of the projection operator. In figure 4.3 we examine 8 different meshes consisted
of triangles and quadrilaterals. We start by projecting sin(-Kx)sin{-Ky) onto a square
domain covered with standard elements. Figures 4.4 and 4.5 show results for the modal
and mixed bases that demonstrate exponential convergence is achieved. Subsequently,
we make the elements covering the domain progressively more skew in the meshes B-
H. In each case we see tha t exponential convergence is achieved, even when one of the
triangular elements has a minimum angle of about 10-3 degrees. Hence, the accuracy
of the method is extremely robust to badly shaped elements. Also, we note that
the similarity of the convergence curves demonstrates that the rate of exponential
convergence is unaffected by the skewing.
The independence of the rate of skewing might be expected based on our previous
observations about the accuracy of projection using the orthogonal basis. We saw
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
74that the coefficients in the expansion of an infinitely smooth function are bounded
by exponentially decaying functions of the polynomial order which are scaled by the
Jacobian of the geometrical mapping. So we infer tha t as an element becomes more
skew the energy in each mode should reduce in proportion to the area of the element.
4.3 Convective Operator
We now consider the two-dimensional linear convective equation for u(x, y, t):
This has been formulated for triangular elements in [20], and we note that using
quadrilateral elements requires very few changes to this method. We consider the
weak form of this equation in £lk, th a t is:
Find u 6 such th a t for all w G H l (£i)
( du du du \ , . ,(a t - a d i ~ % ’w) = 0 ’
Following a Galerkin formulation so tha t the trial and test spaces are spanned by the
same basis we obtain:
dx J V dy
We now define two local operators Bk and Lk{6):
ZLfi
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
75
B k = (<£*,<&)
Lk{0) =COS(0) i t
L ' + s i< e) U d- § -
where we have normalized the convection velocity appropriately so th a t its direction
6 is defined by the scalars a, b. Based on these definitions we now construct two new
operators for the entire domain:
and similarly:
m =
L 1 0 ... 0
0 L i . . . 0
0 0 ... L k
B =
B i 0 . . . 0
0 B 2 ... 0
0 0 ... B k
We can now assemble these elemental operators into a global operator, by means of
the Z operator [40] tha t assembles the local coefficients into the global coefficients
and ensures C° continuity.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
76
Element 2
Element 1
Element 2
Element 1
Local Numbering Global Numbering
Figure 4 .6 : Illustration of local and global numbering for a domain co n ta in ing one quadrilateral and one triangular elements. Here the expansion order is N = 3, and we only show the boundary modes.
To illustrate this global assembly procedure, we consider a global domain made
up of two elements as shown in figure 4.6. The expansion order shown here is N = 3
which means there are six boundary modes on the triangle, and eight boundary modes
on the quadrilateral.
The to ta l number of local degrees of freedom is therefore Niocai = 14. Since three
modes meet along the connecting edge the number of global degrees of freedom is
eleven and so for this case Z is a 14 x 11 matrix as shown in figure 4.7.
The superscripts denote the local or global nodal number and the subscripts denote
the element number. The absolute column sum gives the multiplicity of a mode and
we see th a t columns 5,6 and 7 all have a multiplicity of 2. We also note that the
absolute row sum is always 1 since there is only one value of each local mode. It is
also possible to have a (—1) entry if we have two elements where: (1) edge 1 meets
edge 1, (2) edge 1 meets edge 2, (3) edge 2 meets edge 2, (4) edge 3 meets edge 3, (5)
edge 3 meets edge 4, or (6) edge 4 meets edge 4, then the local co-ordinates at the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
77
1t-H
t
1
“ 3 1 u 9
u \ 1 « 2
1 U9
1 <
U j 1
u l = 1 u f
... u 97
“ 1 1 u 9
« 2 1 u f
1 f.Su 10
u \ 1 1~l9 _ U11
*5 1
. “ 6 . 1
Figure 4.7: Z-matrix map from global to local degrees of freedom
edges are in opposite directions. If we are using a hierarchical basis then we need to
negate odd modes along one of the edges.
Having defined the assembly operation, we can now write the global system as:
du9— = G(SK ,
where G(6) = (Z tB Z )~ 1(Z tL(9)Z).
The behaviour of this operator is very important with regards to determining the
maximum time step we will be able to use for any problem which involves an explicit
treatment of convective contributions.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
784.3.1 A ccu racy o f th e C onvective O p erator
We tested the accuracy of the Galerkin convective operator using a third-order Adams-
Bashforth temporal scheme, and a periodic domain as shown in figure 4.8. We started
with the initial condition u = sm(7rcos(7ra;)) and examined the error a t t = 0
and t = 2. The convection velocity was constant. We chose a time step small enough
so that the time-stepping error is small compared to the initial projection error for
N < 16. We see th a t exponential convergence is m aintained after one time period.
10 ' ' Error at t=0 Error at t=2
L.10'
.<10
10'
10'
Expansion Order
Figure 4.8: Exponential accuracy is achieved for the wave equation with u = sin(7rcos(7rx)) as initial condition.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
794 .3 .2 S p ectru m o f th e G alerkin 2D C on vective O perator
We can now examine the behaviour of this operator by examining its eigenspectrum.
The distribution of the spectral radius, p{9), shows us the level of directional inhomo
geneity of wave frequency supported in a given domain. We first consider a periodic
box that is discretized with essentially standard elements, and then we examine how
deforming these elements within the box affects wave propagation.
Ca)
Figure 4.9: Spectral radius of the weak convective operator on a periodic domain, N = 12.
In figure 4.9 we show three discretizations of the periodic box. The first one (a)
employs only quadrilaterals, the second one (b) a mix of quadrilaterals and triangles,
and the last one (c) only triangles. G{9) was constructed using a 12th order expansion
and the spectral radius of the G{9) for each mesh is shown as a function of 9. In the
upper right quadrant we see tha t the spectral radii are very similar for all three cases,
but in the lower right quadrant we see that there is a marked difference between the
spectral radii of the quadrilateral mesh and the triangle mesh with the hybrid mesh
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
80between these two cases.
Theoretically, we do not have to consider the spectral radius of the mixed basis, as
we have shown that since it is numerically similar to the modal basis it will share the
same spectral properties for linear operators. However, we did run this test for the
quadrilaterals using a nodal basis and obtained the same spectral radius to machine
precision, confirming the theory.
i121111101
<r
Mesh (a) Mesh (b) Mesh (c)
15105Expansion Order
Figure 4.10: Growth of the spectral radius of the Galerkin convective operator with expansion order.
So far we have examined the spatial variation at a given expansion order. In figure
4.10 we now demonstrate tha t supfl p{G{9)) grows as 0 ( N 2) for the modal basis used
on all three meshes, and again we note that the mixed basis has exactly the same
property to machine precision.
We have an exact fit for the numerical spectral envelope of the Galerkin convective
operator on a periodic square domain discretized with regular quadrilaterals. This
can be represented as (see also figure 4.11):
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
81
■°s?
OS
41 •
Figure 4.11: pa(&) is an exact fit for the spectral radius of the Galerkin convective operator on a periodic domain tiled with regular quadrilaterals and with N = 12.
i>’ -.1
.10 -OJ 00 OJ to
Figure 4.12: Pa(#) is a close fit for the spectral radius of the Galerkin convective operator on a periodic domain tiled with triangles and with N = 12.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
82
(c)
Figure 4.13: Spectral radius of the Galerkin convective operator on a periodic domain discretized with non-regular elements.
Pn{Q) =
\/2/?n (O)szn(0 + f ), if 6 > 0
y/2pa(Q) sin{6 + ^ ) , otherwise.
We also have an approximate fit for the numerical spectral envelope of the Galerkin
convective operator on a periodic square domain discretized with regular triangles.
This is described by (see also figure 4.12):
Pa (0 )\/2(pA(O)sm(0 -1- f ) — sin(2Q)), if 9 > 0
y/2(p^(Q)sin(9 + + ( p a ( 0 ) - pa(O))sin{20)), otherwise
We have not so far indicated how the operator G{6) depends on the skewness
of the physical elements. We will now consider the same periodic box discretized
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
83with deformed elements, so tha t these mappings are not simply scaled identities. In
figure 4.13 we have created deformed elements by simply shifting the vertex in the
middle of the quadrilateral and triangle meshes we have just considered. In addition,
we created a new triangle mesh by taking the Delaunay triangulization of the given
vertices. This time we see th a t the deformation of the triangle mesh has increased the
spectral inhomogeneity of the operator, but th a t this can be ameliorated by choosing
the Delaunay triangulization.
4.3 .3 T h ree-D im en sion a l G alerkin C onvective O perator Spec-
trum s
We will now consider the same problem posed in three-dimensions. Each element
type will be shown to have different spectral properties. The three-dimensional linear
convective equation for u ( x , y, z, t ) is:
d u (x ,y ,z , t ) _ du . .d u . . .d u . .d u— a t-----+ L u s a t + a{x' y ' z ) d i + b(x ’v ' z ) a i + c(x' v ' z )T z = ° '
We consider the variational form of this equation in f2fc, that is:
Find u G H l {Vt) such th a t for all w € H l (Q)
( du du ,d u \ . u r r l .9 t + a d i + % + cd-z'wj = 0 '
Applying the Galerkin formulation so that the trial and test spaces are spanned by
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
84the same basis we obtain:
(^ni 0m)dumdt { ^ n,a d x ) + (^ " ’6 d y ) + (? n''
90m : dz
UT
We now define two local operators Bk and LkiO):
B k = (0E.0&)
Lk(0) = cos(0)cos{9) (<#;, + cos(<t>)sin(0) + sm(0)9y dz
where we have normalized the convection velocity appropriately so that its direction
9 is defined by the scalars a, b.
Based on these definitions we now construct two new operators for the entire
domain:
L(9) =
L x 0 ... 0
0 Lo ... 0
0 0 ... Lfc
and similarly:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
85
B =
B l 0 ... 0
0 Bo ... 0
0 0 ... B K
As in the two-dimensional case (section 4.3) we assemble the boundary degrees
of the freedom to make the operator act continuously across the element interfaces.
This time the Z operator maps the globally numbered modes to the locally numbered
vertex, edge and face modes.
4.3 .4 S p ectru m o f th e G alerk in 3D C on vective O perator
As before we examine the spectral radius of the convective operator as a function of
the direction of the wave vector. This is now a vector in 3-space so the variation of
the spectral radius is a two-dimensional surface.
In three-dimensions we have four element types to consider. The observation we
wish to make is th a t the magnitude of the spectral radius of the Galerkin convective
operator scales as N 2 and the directional variation is a function of the element length
in a given direction. This “length” can be defined as the minimum length between a
vertex and an opposing face in the given direction.
We first consider a box discretized with tetrahedra. The spectral radius surface is
shown in 4.14. There is clearly a drop in support at the top rightm ost octant and the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
86bottom leftmost octant. It is evident tha t the elements all have their longest edges
aligned in the direction between these octants.
Figure 4.14: Left: Exploded spectral element mesh, Right: Variation of the spectral radius of the convective operator in a periodic box discretized with six tetrahedra and fourth-order expansion.
Next we consider a box discretized with pyramids. The spectral radius surface is
shown in 4.15. The mesh has a six-way symmetry that is reflected in the symmetry
of the surface. The compression of the distance from apex to base causes a slight
increase in resolution in that direction.
Next we consider a non-unit box discretized with prisms. The spectral radius
surface is shown in 4.16. This is a simple two element mesh showing the effect of
compressing the uniform direction of the prisms. The surface shows overresolution in
the vertical direction compared to the horizontal directions. If we scale the vertical
direction to match the horizontal lengths the we would see tha t the surface is quite
uniform except for underresolution in the direction aligned with the longest triangular
face edges.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
87
k .
-5
. * p £ 2 * « i *»• ‘iV iv iX
"« * '
{ j r - . t a i i *--4 •
. t c ,
Figure 4.15: Left: Exploded spectral element mesh, Right: Variation of the spectral radius of the convective operator in a periodic box discretized with six pyramids and fourth-order expansion.
k '
Figure 4.16: Left: Exploded spectral element mesh, Right: Variation of the spectral radius of the convective operator in a periodic box discretized with two prisms and fourth-order expansion.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
88Lastly we repeated the experiment with a periodic box tiled with eight hexahedra
shown in figure 4.17. The surface of the spectral radius reflects the symmtries and the
same form of over resolution in the diagonal directions as we saw in the quadrilateral
mesh case.
Figure 4.17: Left: Exploded spectral element mesh, Right: Variation of the spectral radius of the convective operator in a periodic box discretized with eight hexahedra and fourth-order expansion.
4.4 Diffusion Operator
We now consider the two-dimensional elliptic Helmholtz equation:
Z
(v 2 - A ) u = f , A > 0.
Again using the Galerkin formulation and integrating by parts we obtain:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
We will define a new set of K operators:89
L t = [ ( V ^ ,V ^ ) + A ( A n,A.)]
and a new set of K vectors:
Then repeating the process to assemble the weak convective operator we can assemble
the subdomain operators into a global operation to obtain:
u 9 = (Z tL Z )~ 1Z tF,
where the bold letters denote vectors.
4.4.1 T h e Schur C om p lem ent M eth o d o f Inverting G lobal
O perators
We are left with a m atrix inversion problem. The m atrix A = Z lL Z has to be
inverted. The Z m atrix groups together the degrees of freedom for the boundary-
mode shapes for all the elements. The remaining interior degrees of freedom are
collected in one group per element. The m atrix now has the form shown in figure
4.18.
We can take advantage of the bandedness of the A boundary-boundary matrix,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
90Boundary d.o.f. Interior djj.f.
A B
Bt C
A = B =
C =
Figure 4.18: Form of the global operator matrix. Notice the sparsity of A and the block nature of B and C.
and the decoupled C matrix. The system can be rewritten as:
A - B C ~ lB T 0 06 f b - B C - l f {
B T C 0i f i
where 0& are an array of the unknown global boundary coefficients and 0,- are the
unknown local interior coefficients.
4.4 .2 O p tim al N um bering for th e Schur C om p lem en t B ound
ary M atrix
To solve the global matrix system the boundary system is solved first . This is done
by using the Schur complement. A direct solve on this system involves processing
a large da ta structure. Here we present a method for applying the Cuthill-McKee
algorithm [38] to reduce the bandwidth of the boundary solve matrix.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
91Although the boundary system matrix is large it is sparse. By reordering the
global numbering of the unknowns we take advantage of the sparcity and pack the
data into a band along the diagonal of the matrix. We define the bandwidth of
the m atrix as the width of this band. The bandwidth is equivalent to the largest
elemental bandwidth taken of all the elements.The elemental bandwidth is defined as
the largest difference between the global identity of the unknow ns in a single element.
All the unknowns in a single edge (or face) are grouped together and numbered
consecutively because they will inevitably be closely associated in an optimal renum
bering. We define an approximate bandwidth as the maximum difference between
the global numbering of any two facets in an element. M inim iz in g the approximate
elemental bandwidth over all the elements is almost equivalent to the original mini
mization problem.
We start this re-ordering at a facet on the boundary of the region and consider all
the elements in which it lies. We order all the facets in these elements in ascending
order of their degree. The degree of a node is defined as the number of facets with
which it shares an element. The new facets are collectively defined as a layer. We
apply this procedure to all the facets in the layer, in their new order, to form a new
layer. We order the facets in the new layer in order of their degree. We apply this
ordering repeatedly to create a sequence of layers that span the mesh.
Because any element will contain facets from two layers the resulting bandwidth is
equivalent to the maximum number of facets in any single layer. We choose the initial
facet so that each of the layers contains the least number of elements possible. Thus,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
92we start the algorithm on the boundary to restrict the directions tha t the layers can
expand. Because we do not know which boundary facet will give us the best result,
we start the algorithm from a number of the boundary facets in the order of their
degree. This number can be set for speed of the overall algorithm.
The described re-ordering leads to a distinct structure in the final boundary solve
matrix. The banded region of the matrix is formed from intersecting “squares” that
show how the different layers formed by the “wave-front” interact.
i Ui
t' k®' i t*■ n' IB
S g i lBfiSSS?7
ii!
Figure 4.19: Here we show the domain that we solved the Helmholtz equation in the following forcing function, / = — (3.47T2) sin (7r.a:)cos(7ry) cos (0.2ir.z) and boundary conditions u = sin (7 r .a :)c o s (7 r .y ) cos (0.27r..z), with periodic boundary conditions at the spiral ends. We see a dram atic bandwidth reduction because of the aspect ratio of the mesh and the exponential decay in the error with increasing expansion order.
We could have performed an O(Nbl) search where N b is the number of unknowns
but this would be prohibitive in all but the simplest meshes.
4.4.3 N u m erica l R esu lts
In this section we aim to examine the effect on computational efficiency of reducing
the bandwidth of the Schur complement matrix. Both optimized and unoptimized or-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
93
p Bandwidth User solve time in secondsinitial final unbanded direct banded direct iterative
3 361 59 0.53 0.50 0.824 1501 244 1.14 0.98 2.415 3441 558 2.59 1.94 6.196 6181 1001 5.7 3.70 17.47 9721 1573 10.95 6.08 39.358 14061 2274 20.45 10.37 76.099 19201 3104 — 16.77 141.66
Table 4.2: Timings for the Schur complement obtained on a single node SGI R8000.
derings have been implemented in the spectral element code M e k T o r . We present
two examples: the first is a solution to a scalar elliptic Helmholtz equation, while the
second is a solution to incompressible Navier-Stokes equations. In order to obtain a
sensible measure of this effect we consider the CPU time for the banded direct solve
and the iterative solve of the boundary system resulting from the Helmholtz equation
( V 2u — u = f ) against the CPU time taken by J ^ e k “T o l t to perform a direct solve
on the unbanded boundary system. However, because of memory requirements we
run into problems w ith the unbanded system becoming too large to run while the
other methods are still very manageable.
We have performed the optimization analysis on the mesh inside a spiralling pipe
as shown in figure 4.20. In this case we solved the Helmholtz equation with a forcing
function designed so th a t we could measure the numerical error. In each of the three
methods the L<i and the errors were identical for a given expansion order.
In table 4.2 we have compared the timings for the backsolve routine only for the
unoptimized and optimized direct solves versus a diagonally preconditioned conjugate
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
94gradient iterative solve. The improvement in CPU time for the banded computations
compared with the unbanded computations is substantial, where an improvement of
an order of magnitude is obtained compared to the iterative solve. The preconditioner
used in the iterative solve is based on the diagonal; in current work we are using the
hierarchies of the expansion basis to obtain more effective preconditioners. The CPU
savings reported here are coupled with an extreme reduction in storage requirements
making this method quite competitive with the iterative method. Of course the
iterative method wins for very high orders p due to memory restriction, but for these
kinds of orders the banded method looks superior given a reasonable memory capacity.
Krror
864
Expansion ordar
Figure 4.20: The exponential decay in the error with increasing expansion order for the Helmholtz equation solved on the spiral domain.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
95
IJ
is i« a s
I
rtf’
(a) (b)
(c) <d)"H
Figure 4.21: Convergence test for th e Helmholtz equation using quadrilaterals and triangles, with Dirichlet boundary conditions. The exact solution is u = sin(irx)cos(7ry) and forcing function / = —(A + 2'K2)s in ( t t x ) c o s (jry).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
96
4 .4 .4 C onvergence for th e G alerk in H elm h oltz O perator
In figure 4.21 we demonstrate convergence to the exact solution with p-refinement
and /i-refinement for the Helmholtz equation with A = 1; the exact solution is:
u = sin(ncos(nr2))
and forcing function / = —(A + 47r4r 2szn(7rr2)2)sin(7r(cos(7rr2))) — 4ir2 (nr2 cos (irr2) -f-
sin(nr2))cos(ncos(7rr2)), where r 2 = x 2 + y 2 for Dirichlet boundary conditions.
10“
Iff*26 61 36 41 4 6 51 S 6 6116 2111
zo
to
OS
ao
1000
Figure 4.22: Convergence test for the Helmholtz, using a triangle and a quadrilateral, with Dirichlet boundary conditions: u = sin(ncos(nr2)) .
In figure 4.22 we show p — type convergence for a more complicated exact solution.
This example demonstrates that the method is stable to a t least N = 64.
C onvergence in 3D H y b rid E lem en ts
We constructed a blocky fish shape domain to test the Helmholtz solver with all the
three-dimensional element types connected together. In figure 4.23 we show th a t the
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
97
solution converges to the correct solution for the Helmholtz problem with increasing
N.
Pyramid
Hexahedron
l<r
10°
• 10“
10-
10'7 r
Expansion Order
X u.
Figure 4.23: Convergence for the Helmholtz problem on a mesh of hybrid elements. Dirichlet boundary conditions u = sin(x)sin(y)sin(z) and A = 10000.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
98Convergence in 3D Skew Elem ents
We have shown in chapter 3 that the approximation properties of the orthogonal
bases should not be adversely affected by the shape of triangles or tetrahedra. We
constructed a twelve element tetrahedral mesh, shown in figure 4.24, for a box. All
the tetrahedra share a vertex at the center or the box. For the first test of the
Helmholtz solver this vertex was at the center of the box. We then moved the vertex
progressively closer to one of the comers of the box and ran the solver for each
position. In figure 4.24 we show that the convergence of the L error norm was only
minim ally effected by the element skewing. This is very good news because we can
use unstructured meshes without worrying too much about mesh quality for accuracy.
We should however still be careful about using elements with small angles because we
have already shown tha t these elements have over-resolution in their shortest vertex
to edge direction. This can lead to a very restrictive Courant-Friedrichs-Lax (CFL)
condition in wave propogation problems.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
99
L _ E r r o r . C e n te r V e r te x = < 0 ,0 ,0 )
L _ E r r o r . C e n te r V e r te x = < 0 .4 ,0 .4 ,0 .4 )
E rr o r , C e n te r V e r t e x = ( 0 .8 ,0 .8 .0 .8 )
L „ E r r o r . C e n te r V e r t e x = ( 0 .9 5 .0 .9 5 ,0 .9 5 )
L . E rr o r . C e n te r V e n e x = < 0 .9 9 .0 .9 9 ,0 .9 9 )
5 10Expansion Order
1 5
Figure 4.24: Convergence is guaranteed for the Helmholtz problem even on very skewed tetrahedra. Dirichlet boundary conditions u = sm (^ )sz n (^ )sm (^ p ) and A = 1.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter 5
Incom pressible Flow Sim ulations
In this chapter we show the results from applying the hybrid grid technology to solve
the Navier-Stokes equations. We modified an existing version of the N e t t T o t r code
[30] to allow the use of quadrilaterals in two-dimensions and hexahedra, prisms, and
pyramids in three-dimensions in addition to the original support for triangles and
tetrahedra. The solvers needed minor modifications due to the general nature of the
original code and the compatibility of the new element representation with the older
C-based representation. We tested the new version of A T e k . 'T otr on simple physical
models in order to verify the robustness of interfacing different element types together.
We then extend the algorithm to be able to solve the Navier-Stokes equations in
moving domains using the ALE formulation. We show tha t the numerical scheme is
accurate for a moving domain test problem.
100
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1015.1 Incompressible Navier-Stokes Equation
5.1.1 F orm ulation
The two-dimensional incompressible Navier-Stokes equations are:
-ST + (v • V )v = —Vp 4- i/V2v + F, at
V • v = 0,
where v denotes the velocity of the fluid with components v = \u(x, y, t ), v(x, y, t)]T
in the x and y directions; p(x, y, t ) is the pressure; F (x, y, t) is a forcing function and
v is the kinematic viscosity. To discretize these equations in time we use a high-order
splitting scheme [39]. A variation of this scheme using the spatial operators we have
already described is explained and demonstrated in [30].
5.1.2 S um m ary o f Schem e, B ou n d ary C onditions and Im ple
m en ta tion
The splitting scheme involves the following four substeps:
v = ^ ocq-vn- q + /3ffN (v"“») + F n+l), (5.1)q = 0 9 = 0
—/£ 1/?,N(v"-'0 - v ^ 13,[V x (V x v -« )] (5.2)g=0 9=0
dpn+l= n
dn
V p = V - ( A ) , (5.3)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
V2v n+1 - ————v n" 1 = ---- — v - A £Vp"+1.i/At i/At
102
(5.4)
This has been implemented in the J\fs k T ctr code [30] for triangular and tetrahedral
elements and in PRISM [40] for quadrilateral (nodal) elements. M erC T a r has now
been recoded in an object oriented way so th a t it now handles all the element types
we have described in chapter 2. The first step we call the non-linear step. We assume
that the velocity fields axe stored in terms of coefficients of the shape functions we
outlined in 3. We then evaluate the velocity fields at quadrature points for each
elements using a transform based on tensor product summation of the modes; this
has been detailed in [30]. We now have a “physical” representation of the velocity at
a set of points. For the non-linear step we take the spatial derivatives of these fields
at the quadrature points and form the the non-linear convection term s by multiplying
the velocity and its derivatives together at these points. We then take the current
values of the non-linear terms and either one, two or three non-linear terms from
previous times steps and add them together w ith a set of coefficients detailed in [39].
We now have the intermediate velocity as a field at the quadrature points.
For the second step we go through the list of boundary conditions for the elements
and calculate the Neumann boundary condition for pressure using the compatability
condition from equation 5.3.
For the th ird step we take the divergence of the intermediate velocity, using the
same spatial element-wise derivative matrices as before. We then take the inner
product over each element with that element’s basis functions. We now have a set
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
103of coefficients for the right-hand-side of the pressure equation. Next we calculate
the pressure field using the implicit solver we outlined in chapter 4, using the Schur
complement of the global discrete Laplace operator.
For the last step we need to evaluate the pressure field at the quadrature points
and then take its gradient also. From this we calculate the right-hand-side for the
velocity solves. We take the inner product of the righthand side with the set of
elemental basis functions. Again we use the implicit solvers to invert the Helmholtz
equations for the velocity.
5.1 .3 W annier F low
The first test we consider is Stokes flow past a rotating circular cylinder next to a
moving wall. The exact solution due to Wannier [41] allows us to evaluate the error
in a domain involving curvilinear elements. The exact solution can be written:
u(x, y i = y + d) 2(A + F Vl) r, K x l('
b r o ~ ~ k [ s + 2 y i ~
2yi(s + y i)(s + yQ ATi
v { x , y i = y + d) = (A + F yx){K2 - K x) -2I 1A 2
2 y i ( s - y i ) ( s - y i ) l _ DK 2 J
rxs Tf \ ^ B x y x{s + yx) 2Cxyx(s - yx)
- D
where we define:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
104
D = —U, F = Uln (T )’
K i = x 2 + ( s + y i)2, k2 = x 2 + (s — 7/i)2, s2 = d 2 + R 2, F —d + s d — s'
Here we have used a cylinder of radius R = 0.25 which is a distance d = 0.5 from the
moving wall. The wall is moving with a velocity of U = 1 and the cylinder is rotating
in a counter clockwise sense with an angular velocity of uj = 2. The domain was split
into 65 elements and the discretized domain is shown in figure 5.1.
to 4
tff4TO*
Iff* -
i<r'to*nr*
to"*
L.
u
tj
u :
-a*
11 13 IS 17
gxpmsioo O rder
Figure 5.1: Discretized solution domain for the Wannier-Stokes flow using 68 elements. Left: Graph of exponential spatial convergence, Top Right: Hybrid spectral element mesh, Bottom Right: Streamlines of the steady state Stokes flow.
The steady state solution with Dirichlet boundary conditions using an expansion
basis of L = 11 is shown in figure 5.1. This figure shows iso-contours of velocity,
streamlines of the steady flow and the exponential convergence to the exact solution
with increasing order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
5 .1 .4 K ovasznay F low105
1.5
0.75
- 0.1 0.5X
0.75
0.25
-0.I0.5
x
Figure 5.2: Steady state solution for the Kovasznay flow a t Reynolds number Re = 40 using a discretization using K = 24 elements. Left: The spectral element mesh used, Right: Steady state streamlines.
The first Navier-Stokes solution we shall consider is the Kovasznay flow. This is a
laminar flow behind a two-dimensional grid, the exact solution of which is due to
Kovasznay [42]. This solution can be written as a function of Reynolds number Re
in the form:
u ( x ,y ) = 1 —eAlcos(27ry)
v (x ,y ) = ^ - e Alsin(27ry)Z7T
where
x Re (R e 2 A 2\ *~ T ~ ( “ i " )
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
106
Parameter ValueDimension 2d
Re 40A t 0.0001
N-Range 5 to 11K th 12
HQuad 12Method Galerkin
Table 5.1: Simulation parameters for the Kovasznay flow.
Using the exact solution as Dirichlet boundary conditions, a steady state solution
was obtained using the discretization shown in figure 5.2. Also shown in this figure
are the steady state streamlines and iso-contours of the rr-component velocity a t a
Reynolds number of Re = 40 using an expansion basis of N = 7.
Using the exact solution allows us to calculate the error, in the L00 and H i norms,
with expansion order as is shown in figure 5.3. A summary of the simulation param
eters is given in 5.1.
We repeated this experiment with a three-dimensional mesh. Again in 5.4 we see
tha t as we increase the expansion order we obtain exponentially increasing accuracy.
5.1 .5 C ylinder F low
Flow past a cylinder provides a good way to verify the M en.'Totr code. For R e > 40
vortex shedding occurs a t the cylinder and a von Karman street of vortices forms in
the wake of the cylinder. This shedding process causes the forces on the cylinder to
oscillate with a distinct frequency. This is known as the Strouhal frequency (S t). The
Strouhal frequency will be used as a measure to compare the results from the hybrid
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
107
L_ Error in H, Error in L_ Error in H, Error in
10'
10 '
10'
10'
6 115 7 8 9 10
Expansion Order
Figure 5.3: Convergence in the L ^ and H\ norms as a function of expansion order for the steady state Kovasznay flow at a Reynolds number Re = 40.
A /s k T a r code and other codes. For Reynolds numbers up to approximately 190
the flow structures remains two-dimensional. Above this number three-dimensional
instabilities occur, causing the two-dimensional approximation to be increasingly in
accurate for higher Re. Thus, we will consider a range of Re up to 250 in order to
investigate how difficult it will be to resolve high gradient fields.
For this test of the hybrid incompressible code we consider two-dimensional flow
past a circular cylinder. The cylinder has unit diameter, and the domain surrounding
the cylinder is a rectangle [—22,69] x [—22,22]. Uniform velocity boundaries are used
a t the inflow, upper and lower boundaries. Zero Neumann boundary conditions are
used for velocity and the pressure is set to zero at the outflow. Figure 5.5 shows the
domain, mesh and boundary conditions.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
108
10 ' '
icr
10
1CTui ^
w 10-6
10"
1CT
1 0 '
10- ’
J L . 1 1 ' 1 ' 1 1 ' 1 ' i L10 12
Expansion Order
Figure 5.4: Top Left: Three-dimensional spectral element mesh used to solve Kovasznay flow on, Top Right: Iso-contours of streamwise component of velocity, Bottom: Convergence to exact solution with increasing order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
109
Parameter ValueDimension 2d
Re 100 to 250A t 0.002
N-Range 10 and 12K th 448
R-Quad 332Method Galerkin
Table 5.2: Simulation parameters for the incompressible cylinder flow simulation.
Resolution Re=100 Re=150 Re=200 Re=250Hybrid N e k T a r K = 780, N = 10
K = 780, N = 120.16590.1662
0.18510.1853 0.1972 0.2054
M s n 'T a r [30]) K = 173, N = 11 0.1667 NA 0.1978 NAPrism [12] 14000 dof 0.1664 0.1854 0.1969 0.2051
Table 5.3: Variation of Strouhal frequency with Reynolds number for incompressible flow past a two-dimensional cylinder
The mesh has 332 quadrilaterals and 448 triangles. It was designed to direct res
olution around the cylinder, have regular resolution in the wake behind the cylinder,
and block out with large elements to push the farfield boundaries away from the
cylinder. We ran the simulation at N = 10 for Re = 50,100,150,200 and 250.
In the following table we show good agreement between Strouhal numbers obtained
by Hybrid M e k 'T a r , M e k T a r [30] and Prism [40]. A summary of the simulation
parameters is given in table 5.2.
In figures 5.6 and 5.7 we show iso-contours of vorticity of the wake region of the
cylinder at Re = 50,100,150,200 and 250. We can see that at Re = 50,100 and
150 the vorticity is smooth and the vortex structure is well resolved. At R e = 200
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
110
Resolution Re=100 Re=150 Re=200 Re=250Hybrid M s k T a r K = 780, N = 12 1.34468 1.33136 1.34473 1.36339
Prism [12] 14000 dof 1.35000 1.33336 1.34116 1.35769
Table 5.4: Variation of C& with Reynolds number for incompressible flow past a two-dimensional cylinder
and more so a t Re = 250 we see th a t the bulk shape of the vortices is correct but
the features are a little unresolved. There are also streaks of spurious vorticity at
Re = 250. These effects will be reduced by increasing resolution.
M e n T a r has been taken to higher Reynolds number flows past a cylinder with
the two-dimensional module and with the Fourier module I wrote based on the model
in [40]. The Fourier module uses two-dimensional planes connected with a homoge
neous third direction for which we use a Fourier expansion. The results from these
simulations have been published by Ma Xia [45] and Evangelinos et al. [46]. The
numerical method was adapted from [40] to include both triangles and quadrilaterals
in the two-dimensional planes and allow variable expansion order from element to
element. The Fourier module is also being adapted by Evangelinos [47] to include a
time dependent spatial m apping in the third direction. T hat work is an adaption of
the work by [48] where only nodal quadrilaterals were used in the two-dimensional
planes.
5.1.6 T h e R o b o tu n a
So far we have demonstrated th a t Ss/ s k 'T a r can correctly simulate simple flows in
relatively simple geometries i.e. flow past a cylinder or flow in a box. Now we push
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
I l l
**=(1,0)
20
p = 0
-10
•206 0-20 -10 0 20 3 0 4 0 5010
ao(l,0)
\ / \ / \ / \ /___ ------------1 >t----------
- / \ / \ / \ / \ iiiiiii
i ; i(
D * ^ V \ Ai ii i
i
i i I !!
I /,\ 1 A l A 1 XX . , . .L0 5 1 0
X
Figure 5.5: Domain and mesh for the two-dimensional incompressible flow past a circular cylinder simulations.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
112
Re 50
o
10
Re 100
' '
Re 150
■SSe'i ®8i
Figure 5.6: Instantaneous contours of vorticity for incompressible flow past a two- dimensional cylinder at Reynolds numbers Re = 50,100 and 150.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure 5.7: Instantaneous contours of vorticity for incompressible flow past a two- dimensional cylinder at Reynolds numbers Re = 200 and 250.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
114the envelope by considering the flow past a Robotuna. Triantafyllou et al. at the
towtank facility at MIT have constructed a robotic tuna tha t can swim up and down
the towtank. They kindly provided the surface definition of the Robotuna device.
W ith the help of PhD student Mike Kirby [82] and FELISA [3] we constructed a
tetrahedral mesh around the surface definition. Ideally we would have liked to use
a thin layer of prisms around the fish surface, but as a first pass it was instructive
to use an off the shelf tetrahedral mesh generator to produce the mesh. As mixed
element mesh generators come on line to meet demand we expect that this task will
be as straightforward as tetrahedral meshing is today.
We did come across some problems in generating the unstructured mesh. The
actual robotuna is covered by Lycra, so we used a spline to approximate the surface
curvature. However, in doing so we found tha t this isoparametric represention [23]
of the surface caused a small minority of surface elements to become tangled. The
tangling is caused by the curved face projecting through one or more of the other faces
of the tetrahedron. To overcome this problem we identified the tangled elements and
replaced the curvature w ith linear blending of the neighbouring elements’ curvature.
In future work the curvature problems will be removed by the use of prismatic
elements at the surface of the fish. Only extreme deformation of a prism’s triangle
face will cause the prism to become tangled.
In figure 5.8 we show the mesh and surface spline on the surface of the tuna and
in figure 5.9 we show the instantaneous pressure on the surface of the Robotuna. The
high pressure at the nose of the Robotuna and at the front sections of the tail are
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
115as we would expect for a blunt body flow. The pressure on the wider parts of the
body is slightly rough. This may be due to the slight roughness of the surface there
introduced by the surface fixes we have already mentioned.
5.2 ALE Incompressible Navier-Stokes
The Arbitrary Lagrangian Eulerian (ALE) method has been developed by [49, 50, 51,
52, 53] and applied in to quadrilateral spectral elements in [54].
So far we have considered domains that are unchanging in time. The next step
is to allow parts of the domain to move in a time dependent manner. This is not
a trivial generalization because we now have to discretize time dependent operators
as well as time dependent fields. We have made an initial step here, however the
implementation we have used is far from ideal. The principle we wish to demonstrate
here is that the ALE approach using unstructured spectral elements is robust and has
great potential. We must adm it tha t a lot of work (and questions) remains in order
to efficiently implement the time dependent operators. For the moment we reform
the local elemental operators a t every time step. We do this because as yet we do
not have an effective preconditioner for the full implicit solver. By experiment we
have determined that it is more efficient in two-dimensional simulations to use the
fully direct Schur complement method for low (N < 6) simulations and the iterative-
boundary direct-interior Schur complement method [30] for higher order. Bearing in
mind the aforementioned caveats we will demonstrate the possible advantages of using
the ALE method with the spectral element method. Namely, the usual advantages of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
u°!ssi,lUJJdcj
^ N o j d u°!lOn'POJddJ^ JdUMo
W'Mdoo
drJZ :-?ss
^ ^ - ^ Z ^ o J p o q
Op OS
0 T
" ~ * ho
M issio n
118minimal dispersion problems, high accuracy and the novel fact that since the spectral
elements are larger and less numerous than finite element or finite volume elements
they can support larger deformations without becoming entangled. We have also
shown that the triangles and tetrahedra can support deformation without loosing
excessive resolution.
5.2.1 Formulation
We now consider the incompressible Navier-Stokes equations in a tim e-dependent
domain f2(t)
Ui,t + UjUij = —{p5ij)j + v u ij j + f i in 0(£) (5.5)
Ujj = 0 in f2(£), (5-6)
where v is the viscosity and /,- is a body force. We assume for now homogeneous
boundary conditions; specific boundary conditions will be presented in a later section.
We present the variational form of the momentum equation following the derivation
in [54]. To this end, we consider test functions in the space where all functions and
their derivatives are square-integrable, i.e. i l 1 [£!(£)]. Multiplying equation (5.5) by
test functions and integrating by parts we obtain
J V i fa t + UjUij) dx = j f ^ VijipSij - v u i j + Vifi)dx. (5.7)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
119The next step is to define the reference system on which time-differentiation takes
place. This was accomplished in [54] by use of the Reynolds transport theorem and
by using the fact that the test function z/t- is following the material points; therefore,
its time-derivative in th a t reference frame is zero, i.e.
dvi ,-jjr\xp = v i,t + w j viJ = 0
where Wj is a velocity th a t describes the motion of the time-dependent domain Q(t).
Amd X p denotes the m aterial point. The final variational statem ent then becomes
4 f ViUidx + f [viiuj - Wj)uij - ViUiVJjtj\dx = f [vidp5i:j - uvidu^u + Vifi]dxdt Jn(t) 7n(t) 7n(t)
(5.8)
This is the ALE formulation of the momentum equation. It reduces to the familar
Eulerian and Lagrangian form by setting Wj = 0 and Wj = Uj, respectively. However,
Wj can be chosen arbitrarily to minimize the mesh deformation.
5.2.2 T em poral D iscretiza tion
In order to describe in simple terms the solution algorithm, it is convenient to write
equation (5.8) in m atrix form. Thus, we introduce the mass m atrix M = (<£(x), 0(x)),
the derivative m atrix D = (0 ,0X), and the stiffness (Laplacian) m atrix L = <t>x),
where the basis functions <p(pc) are defined in section 3. We also denote the nonlinear
contributions by N (U , W ) to denote dependence on the two velocities, i.e. the flow
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
120velocity field U and the mesh velocity W . Using this notation, equation (5.8) becomes
+ AT<(U, W ) = D j P - LijUj + Ft. (5.9)
and the incompressibility condition (5.6) is
DtUi = 0. (5.10)
The mesh velocity is defined based on the mesh coordinates from
IT = w > - (5'n)
In the above equations we have denoted the fields with capital letters to denote the
discrete version of the continuous field, e.g. ut- £/,•, etc.
Before we proceed with the boundary conditions, we discretize equations (5.9,
5.10) and (5.11) in time using a splitting scheme and a third-order stiffly stable
integration scheme (see reference [55]).
We solve for time step (n+1); first treating explicitly the nonlinear and the mesh
velocity terms:
= ~ Y . ^ N { U n~ \ W n- q) + F n+l9
= E / W 4-9
Ui - 'E .q<xq{M U i)n- qA t
to x r i - n Qa qx r qA t
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
121Next we trea t the elliptic terms implicitly
L P n+1
7o(M U j)n+J- - &j A t
LW \1+1
where we compute the intermediate field from
b i = M U i 4- D j P n+l.
The constants a q, (3q are integration weights and are defined in [55]. The mesh velocity
is in general arbitrary, and it can be specified explicitly or be obtained from a Laplace
equation as done here following [54]. More recent work in [53] suggests a modified
approach where a variable coefficient is used in the Laplacian in order to provide
enhanced smoothing and thus prevent sudden distortions in the mesh.
5.2.3 R ay le igh S lid ing P la te P rob lem
We verified the ALE code using the Rayleigh problem. The domain is a box with a
sliding plate at the lower side. The plate slides horizontally with velocity cos(cdt). The
plate has a periodic length 2. Exact boundary conditions are imposed at y = —0.5,1.5.
The exact solution to the Navier-Stokes equations is:
= vLU?+1
= 0,
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
122
Param eter ValueDimension 2d
Re 40A t 0.01 and 0.001
N-Range 2 to 9K th 50
Method ALE Galerkin
Table 5.5: Simulation parameters for the Rayleigh problem.
u = e vcos(y — ut)
v = 0
p = 0
(y + 0.5)
First we tested convergence to the exact solution using zero velocity initial condi
tions. Figure 5.10 shows that a t N = 5, A t = 0.01 the simulation converges to 10~5
accuracy. Also we show that with A t = 0.001, starting from the exact solution and
checking errors at t = 2, we obtain exponential convergence with increasing N. A
summary of the simulation parameters is given in table 5.5.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
123
1.5
1.25
1
0.75
0.5
>0.25
0-0.25
-0.5
-0.75
N r r f aa S 4 V . j s « a a 5 * '■»
Sliding Plate1 1 ' 1 1 ' 1 1 ' I i i J-L.
-1 -0.5 0x
0.5
56r
Iff’ -SO 100t
10* *.10- -!
10--3
£-j'tff'
Expansion Order
Figure 5.10: Top: Spectral element mesh K th = 50. Lower Left: Time convergence history for zero initial conditions N = 5. Lower Right: Convergence to exact solution with increasing N.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1245.2 .4 D esign o f a M icro-P um p v ia S im u lation
The following section is a summary of a paper subm itted to ASME [56]. The research
was done with Ali Beskok of CFM, Brown University. I t is included to demonstrate
that the ALE component of ft/en'T'ccr has been used as a conceptual design tool.
The goal of this project was to produce a potential design for a pump that can be
fabricated in micro-dimensions.
The performance of the micro-pump is evaluated as a function of the Reynolds
number and the geometric parameters. The volumetric flow rate is shown to increase
as a function of the Reynolds number. However, the efficiency of the micro-pump
decreases with increased Reynolds number, due to the increased leakage effects.
Micro-pump systems, delivering volumetric flow rates in the order of 10-8 ~
10~12m 3/ s can be used in many bio-fluidic, drug delivery, mixing and flow control
applications. Most of the micro-pump systems are actuated by a vibrating membrane
in a chamber with hanging-beam-type (Cantilever beam) inlet and exit micro-valves
[57], [58], [59]. These are uni-directional micro-pumps, since the Cantilever-type
micro-valves only open in a preferred flow direction.
In this study, we present a bi-directional (reversible) micro-pump utilizing a vi
brating membrane and piston-type (moving) inlet and exit valves (See figure 5.11).
Since the inlet and the exit valves are simple micro-pistons, oscillating in between
open and closed positions with a prescribed motion, it is possible to control the per
formance of the micro-pump with control of the micro-valves. The design is flexible,
so tha t the pumping direction can be easily reversed. This design has the advantage of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
125
Micro-Pump0.2L Inlet Valve Exit Valve 0.2L
0.2L ®min — 0.025L Smax — 0-125L
H=0.4L _m__a=0.1 L
Membrane
L[ OutOut
Figure 5.11: Sketch of the micro-pump operating between two micro-channel systems. Inlet and exit valves open and close periodicly with maximum gap of gmax = 0.125L and minimum gap of gmin = 0.025L.
reversibility, yet it can maintain its performance for relatively high Reynolds number
applications.
5.2.5 M icro -P u m p G eom etric S p ecification s
The micro-pump geometry is presented in terms of the length of the membrane “L” ,
in figure 5.11. This allows us to interpret the results using geometric similarity in
determining the optimum pump dimensions.
The micro-pump is ideally placed in between two reservoirs. However, numer
ical simulation of such a system is difficult due to large reservoir size. Any finite
size reservoir would require inflow and outflow numerical boundary conditions, and
these must be carefully imposed, in order to avoid a preferred flow direction in the
micro-pump. This difficulty is overcome by placing the micro-pump in between two
symmetric micro-channel flow systems, where equal amount of liquid flow is main-
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
126tained from top to bottom direction. Therefore, the flow conditions are symmetric
and there is no net flow from one channel to another, when the pump is not actuated.
We verified this by simulations.
MembraneDeflection
o 0.6 0.8 10 .4T T T T
Exit ValveClosedPos.
- EJECTION EJECTIONSUCTION
Inlet ValveOpenPos.
0 02 0 .4 0.6 0.8 1
tcj
Figure 5.12: Top: Deflection of the membrane y(x ,t) = asm(jrx/L) sin(tut) Bottom: Loci of the valve tips y (t) = ± tanh(4cos(cu£))
Oscillation of the membrane with a specified frequency ui and amplitude a excites
the fluid within the micro-pump cavity. For our simulations we have used a = L j 10,
and uj = jj-, where c2 = T / M , T being the uniform tension, and M being the mass per
unit area of the membrane. The first-mode of vibration of the membrane is used to
determine the position (See figure 5.12, top), velocity and the acceleration a function
of time
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
127
y(x ,t) = asin(irx/L) sin(irct/L),
v(x ,t) = (5.13)
(5.12)
(5.14)
The pump geometry is symmetric. Therefore, active pumping can not be achieved
by only actuating the membrane. We have included inlet and exit valves, which open
and close periodicly, in order to break the flow symmetry. The inlet and the exit
valves are located a t the mean height of y0 — 0.325L from the membrane, and the
position of the valves are specified as a function of time
The valve motion is designed to be close to a step function, oscillating between open
valve tips and the top wall of the micro-pump are 0.125L and 0.025L, during open
cycle of the micro-pump is presented in figure 5.12, (bottom). The phase difference
y{t) = ya ± tanh [Acos(irct/L)\. (5.15)
The prescribed velocity of the valve tips are
v (£) = [l — [tanh(4cos(7rc£/L))]2j sin(wct/L).
and closed positions with finite velocity and acceleration. The gaps between the
and closed positions, respectively. The positions of the inlet and exit valves during a
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
between the inlet and the exit valves is 7r.128
5.2.6 E fficien cy A nalysis
The performance of our design is based on the following factors: the membrane length
L , the membrane width W , the pump-cavity height H, the amplitude of vibration of
the membrane a, the frequency of vibrations ui, the minimum valve clearance (the gap
between the closed-valve and the top wall) g, the time-lapse in between the opening
and the closing of the valves (see equation (5.15), and figure 5.12, bottom ) £-1, the
dynamic viscosity of the fluid p , and the fluid density p. There are nine variables
associated w ith the performance of the micro-pump, with-dimensions of length, time
and mass. This corresponds to six non-dimensional variables: j-, and2 _
esjfL- In this study, we have fixed 5 — 0.15a;. The geometric length-scales are set as
H = 0.4L and g = 0.025L. The parameters ^ and and es L are varied.
2The magnitude of the membrane velocity u — uia. Therefore the parameter es L —
^ is the Reynolds number. Since, the ratio of the dynamic viscosity to the fluid
density is the kinematic viscosity v = p./p, the Reynolds number can be simplified as
The volumetric flow rate per channel width (W ) can be calculated by using equa
tion (5.13)
(D r-k— = / v(x, t)dx = — 2accos(irct/L). (5.16)w Jo
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
129The suction stage of the micro-pump happens while — < t < Therefore, average
volumetric flow for a given period T = w~l is
— = —2 a c W c o s ( 7 r c t / L ) A t
4 aL W7T
The average flow rate is
— 4 aLWuj 4 LQ = ---------- = ------W uR e. (5.17)
7r 7r a
This simple analysis indicates that the volumetric flow rate is proportional to the
Reynolds number, the width of the micro-pump membrane W , and the ~ ratio. Our
analysis has assumed no leaks from both the inlet valve during ejection stage and
from the exit valve during the suction stage. Therefore, equation (5.17) gives the
maximum theoretical volumetric flow rate of the micro-pump system. This value will
be used in the next section in determining the efficiency of the micro-pump, while
leakage effects due to the imperfect motion of inlet and exit valves will be considered.
5 .2 .7 N u m erica l S im ulation
Full closure of the inlet and exit valves require annihilation of the elements trapped
in between the valves and the top wall, requiring re-meshing of the computational
domain. Re-meshing is avoided by allowing a gap in between the valves and the top
wall. This gap was g = 0.025L a t closed-valve position.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
130
7 N
Figure 5.13: Spectral element mesh used for the discretization of the micro-pump system at ejection-stage (top, membrane is moving up). The discretization of the flow domain during suction stage is shown at the bottom (membrane is moving down). The bottom figure also shows elemental discretization obtained by 7th order modal expansions used in M en T a r .
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
131
Param eter ValueDimension 2d
Re 0.3, 3 and 30A t 0.0025
N-Range 5,6 and 7K Tri 222
Method ALE Galerkin
Table 5.6: Simulation parameters for the micro-pump simulation.
The mesh used in this study is presented in figure (5.13) at various time instants,
corresponding to suction and ejection stages. The flow domain in represented with
222 triangular (unstructured) spectral elements. Each element utilized T^-order poly
nomial modal expansions. We used successive p-refinements to verify convergence.
The results at 5</l-order and 7^-order modal expansions show that we are resolved
well beyond the scientific accuracy of 1%. A summary of the simulation parameters
is given in table 5.6.2
Re=30i
•S o
-20.6—r~ O.B—ro 0.4—r
R e=3.0
tcj
Figure 5.14: Non-dimensional volumetric flow rate variation with in a period of the micro-pump, as a function of Reynolds number, Re = (a /L = 1/10, a /h = 1/3).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
132In figure 5.14, we present the instantaneous flow rate variation for two different
Reynolds numbers (Re = 3,30). The volumetric flow rate of fluid entering through
the inlet valve is taken as positive and leaving flow rate is taken as negative. The
sum of the two is the rate of change of the control volume due to the oscillation of
the membrane
d V _ ■q = Qin Qout-
The membrane’s motion is periodic. Therefore, the net amount of fluid displaced by
the membrane in a period is zero
j> dV = 0 = J Qindt - J Qoutdt. (5.18)
In other words, Qin = = Q. Numerically integrating the curves under inlet and
exit valves in figure 5.14, we determined the effective flow rate in our micro-pump for
g = 0.025L. The ratio of the numerical values of the flow rate to the maximum flow
rate given by equation (5.17) defines the efficiency (77) of the micro-pump. The results
are presented in table 5.7. It is clearly seen tha t the efficiency of the pump decreases
with increasing the Reynolds number (Re = ^ ) - The average flow rate of the pump
increases with the Reynolds number, as predicted by equation (5.17). This is either
due to the increase in the size of the pump (increase in a), or it is due to increase
in the frequency u . For fabrication of our conceptual design, actual-dimensions of
the micro-pump can be determined by either selecting the admissible amplitude of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
133
Re 0.3 3 30Q fm2]W . * . TrrF.ORRTTdAT. 3.82 x 10"5 3.82 x 10~4 3.82 x 10"3
Q m2W . S . NTTMP.mr.AT. 3.51 x 10"5 3.36 x 10“4 2.86 x 10"3EFFICIENCY (77) 92% 88% 75%
Table 5.7: Volumetric flow rate per unit width W as a function of Reynolds number for closed-position piston-wall gap of g = 0.025L. The value for v ~ 1 x 10-5 for both air and water are used. The mass flow rate is M = pQ
vibration a or the actuation frequency to. For example, if we select a frequency range
1 k H z < to < 100k H z , we can determine the values of the amplitude a and hence,
the rest of the pump-dimensions can be determined.
There is a dominant vortical structure near the inlet valve through out the simu
lations, although its strength varies during the cycle.
The vorticity contours are presented for Re = 30 flow a t various time instants in
figure 5.15. The top figure is a t too = 0.28. This corresponds to the beginning of the
suction stage. S tart-up vortices due to the opening of the inlet valve can be identified
as a vortex pair just a t the top of the inlet valve. The middle figure corresponds to
the end of the suction stage (at too = 0.72). A vortex jet pair is visible in the pump
cavity. The flow pattern at too = 84, corresponding to early ejection stage is given in
the bottom figure. The exit valve has just open, and the start-up vortex due to its
motion is visible at the top of the exit valve. The vortex pair in the pump cavity has
evolved further. The negative vortex is trapped in the pump cavity, and the positive
vortex jet hits the membrane and bounces back, towards the middle of the pump
cavity. Presence of the vortex pair also creates strong vorticity on the membrane. In
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
134
Figure 5.15: Vorticity contours for Re = 30 simulation. Top figure at t u = 0.28, corresponds to the beginning of the suction stage. Start-up vortices due to the motion of the inlet valve can be identified. Middle figure is at t u = 0.72, corresponding to the end of the suction stage. A vortex je t pair is visible in the pump cavity. Bottom figure r u = 84, corresponding to early ejection stage. Further evolution of the vortex jet and the start-up vortex of the exit valve can be identified.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
135
Figure 5.16: Close up of the vorticity contours for Re = 30 simulation a t the left valve (meshes shown on right side). Top: t u = 0.28, corresponds to the beginning of the suction stage. Start-up vortices due to the motion of the inlet valve can be identified. Middle: t u = 0.72, corresponding to the end of the suction stage. A Vortex je t pair is visible in the pump cavity. Bottom: t u = 84, corresponding to early ejection stage. Further evolution of the vortex jet and the start-up vortex of the exit valve can be identified.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
136figure 5.16 we have repeated these time slices but showing a close up of the left valve
of the pum p and the mesh there. When the valve is closed we can see that the mesh
is very tightly packed yet the fields still look convincingly smooth around this area.
After demonstrating the performance of the micro-pump with two-dimensional
simulations, we also tested the design with three-dimensional simulations. We have
extended the two-dimensional geometry of the micro-pump uniformly, using W =
0.4L. We observed boundary layer growth on the side walls, which reduces the flow
rate slightly. Nevertheless, the micro-pump performance is also acceptable in three-
dimensions. The stream-wise velocity contours during the suction and the ejection
stages are presented in figure 5.17.
The algorithm is fast enough to test two-dimensional conceptual design on a work
station. Low Reynolds number (Re < 5) simulations require about 222 triangular
spectral elements with 5th order expansions. The simulation takes about 0.6 CPU
seconds per time step on a 195 MHz Silicon Graphics Onyx2 work station. Three-
dimensional simulations however require large memory and CPU times. A parallel
version of the algorithm is used to test three-dimensional effects on several design
cases. For three-dimensional runs, we have used 444 spectral elements prisms, utiliz
ing 5th order polynomial expansions per element. There are 90 calculation points per
element. The computational domain is divided into 8 sub-domains and the parallel
simulations are performed on the IBM SP2 (thin-2nodes) at Brown University, Center
for Fluid Mechanics. We have observed 18 CPU seconds per time-step for the parallel
run.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Figure 5.17: Three-dimensional micropump simulation Re = 3. Top: An instant during the suction stage. Bottom: An instant during the ejection stage.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 6
Incom pressible V iscous
M agnetohydrodynam ics
The magnetohydrodynamic (MHD) equations are a model for the dynamics of an
overall electrically neutral fluid, called a plasma, tha t is made up of moving charged
particles. These moving particles react to magnetic fields and their relative motion,
as charged particles (i.e. a current), creates magnetic fields too. Thus, the system is
non-linear and a good target for spectral methods. There is interest in the dynamics
of a plasma confined in a toroidal geometry such as a fusion reactor. Clearly, it is
desirable to be able to also consider plasm a in more complicated geometries, thus
spectral elements appear to be an ideal tool for this type of simulation.
The model as stated is actually a two fluid problem with a positively charged fluid
of ions and a negatively charged fluid of electrons. The mass of an electron is very
small compared to the mass of an ion, thus we can represent the system as a single
138
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
139fluid th a t reacts with magnetic fields.
For this section we will assume th a t the plasma is incompressible. This approx
imation greatly reduces the complexity of the system of equations for the model.
Using a simplified model th a t is easily verified provides a good benchmark for the
more complex compressible model th a t we investigate in chapter 8.
The MHD equations have been treated numerically by many including [61] and
[62, 63] using finite differences and [64] using Fourier collocation. There has been some
disagreement about how to deal with the constraint th a t no magnetic monopoles can
exist (i.e. th a t V • B = 0 where B is the magnetic field). However, it has been shown
in [65] th a t the smallest errors in satisfying this constraint can cause a significant
cumulative effect as a simulation progresses. We will test a magnetic streamfunction
formulation for the incompressible equations as this provides a formulation th a t fits
naturally into the set of operators tha t we developed in previous chapters. A simple
modification was needed to make the incompressible Navier-Stokes code handle the
extra field and interactions.
6.1 Formulation
The non-dimensionalized equations for incompressible magnetohydrodynamics can be
expressed as:
< V 1— + v • V v — (V x B) x B = —Vp + — V2vdi
^ + V x ( B x v ) = —J - V x ( V x B )
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
140
Variable Description1 Magnetic resistivityP- Viscosity
O _ PoVa O— u Viscous Lundquist numberSr = ^ Resistive Lundquist number
Alfven wave speed
II
§5
|
Alfven Number
L 0 Characteristic length scaleV0 Characteristic velocityPo Characteristic density
Table 6.1: Parameters for the equations of incompressible MHD.
V - v = 0
V B = 0
where v denotes the velocity of the fluid with components v = [u{x, y, t) ,v(x , y, t)]T
in the x and y directions; p(x, y, t ) is the pressure; and B denotes the magnetic field
of the fluid with components B = [Bx(x, y, t), B y(x, y, £)]T in the x and y directions.
The non-dimensional parameters are shown in table 6.1.
In two-dimensions we used a magnetic streamfunction <j> to represent the magnetic
field:
B = V x (<£k)
This will automatically satisfy the magnetic divergence condition. The two magnetic
fields will now be evolved by the equation:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
6.2 Numerical Scheme
The numerical splitting scheme we chose is:
g=0 9=0
2 JLTl+1 TO'S'r in+1A t A V '
J i - l Jc- 1
v = Y l a ,v n- ? + A ^ /?,Nvv n- q,B n9= 0 9=0
V2v n+1 - ^ 2 ° v n+1 = v - AtVpn+1. A t A t F
B n+1 = V x (<£n+1k)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
where:142
= v x B
N v = —v • V v + (V x B) x B.
The constants a q, Pq and 70 are integration weights and are defined in [55] for the
stiffly stable splitting schemes of first-, second- or third-order depending on the set of
coefficients chosen. The convective terms are treated explicitly and the viscous terms
are treated implicitly. The boundary conditions for the magnetic streamfunction can
be set to 4> = UooV f°r uniform magnetic field at a boundary or <f> = 0 for an infinitely
conducting wall.
6.3 Magnetic Pearson’s Vortex
In order to test the code we tried a simple exact solution to the above equations. It
is a simple extension of the Pearson’s vortex solution for the incompressible Navier-
Stokes equations. We used a periodic box, length 1, and set the initial conditions at
t = 0:
u = — sin(2'Ky)e~ui‘!t2t
v = sin(2TTx)e~‘/A7r2t
B x = —sin(2'Ky)e~v*n2t
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
B y = sin( 2Trx)e-vi*H
4> = (cos(27ry) + cos(2nx))e~t'4w2t
143
We discretized the domain with a four-by-four array of quadrilaterals as shown
in figure 6.1. Also provided are the pressure and magnetic streamfunctions a t time
t = 2. In figure 6.3 we show th a t the error at time t = 2 decays exponentially with
increasing expansion order. The parameters used are shown in table 6.2.
I t H I itiai it h i
ininin0.8
0.7; I t 911uniliill0.6
uni liin it til
0.4
it an nailII H I I t H I
0.1
x
nit ' ; ' '
'O —* . ----.N
0.0515717
>sQ.545.012*44145.0257236
41.0314971
0.12646
00961313
410I1063I410361315410541973•0072263•O09Q32SS-0101393•012646
\ \ :
Figure 6.1: Incompressible magnetic Pearson’s vortex (t = 2, instantaneous fields). Top: Periodic spectral element mesh, Bottom Left: Pressure field, Bottom Right: Magnetic streamfunction.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1 1 , 14 6 8 10 12
Expansion Order
Figure 6.2: Incompressible magnetic Pearson’s vortex. Convergence plot for L2 error in x-component of magnetic field a t time t = 2 versus expansion order.
Parameter ValueDimension 2d
s v 100Sr 100A t 0.001
N-Range 10 to 25K-Quad 64
Method Galerkin
Table 6.2: Simulation parameters for the incompressible MHD Pearson’s vortex simulation.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
145We also used this test case to determine the time accuracy of the splitting scheme.
We fixed the expansion order a t N = 10 and run the simulation to t = 1 for different
values of At. The maximum A t was dictated by the CFL condition. In figure 6.3 we
show that using the first-order coefficients resulted in a unit slope in the log-log plot of
L 2 error versus At. For the second-order coefficients we see second-order convergence.
■ n=10. 1“ order coeff.s ♦ 11= 10, 2“ order coeff.s
10 '
10: ♦
IQ-10'd t
Figure 6.3: Incompressible magnetic Pearson’s vortex. Time accuracy plot for the simulation run with N = 10. Convergence plot for L2 error in x-component of velocity a t time t = 1 versus time step At.
6.4 Orszag-Tang Vortex
The Orszag-Tang vortex is an initial value problem in a square periodic domain
(length L). It has been investigated by [66] and [64] amongst others . It demonstrates
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
146
Param eter ValueDimension 2d
100s r 100A t 0.001
N-Range 10 to 25K Q u a d 64
Method Galerkin
Table 6.3: Simulation parameters for the incompressible MHD Orszag-Tang vortex simulation.
that turbulent scales can result from a coherent initial condition with only two spatial
frequencies. The initial conditions are:
• r2ny\L/
■ ( 2 t x \ v = 52n(—— )L/
B x = -sin (y )
_ . ,4irx.By = szn(— )
An appropriate choice of initial magnetic stream potential is:
L , 2 t t y L f 4 t t x ^
* = ~
As the solution evolves in time we see that the initial vortex splits into two vortices.
The primitive variable fields are shown in figure 6.4 a t t = 1. We see that sharp
gradients build up and it took 64 quadrilateral elements a t N = 25 to resolve these
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
147
Parameter ValueDimension 2d
sv 100Sr 100VA 0.1A t 0.005
N-Range 10K m 460
Method Galerkin
Table 6.4: Simulation parameters for incompressible MHD flow past a cylinder.
features well. We see in figure 6.4 that the vorticity is smooth and well resolved,
and that the velocity divergence is small. The vorticity exhibits complex small scale
structures. Also in figure 6.4 we see tha t the divergence free conditions for the velocity
field is met to 10-4 and the magnetic field is divergence free to 10~8. A summary of
the simulation param eters is given in table 6.3.
6.5 Cylinder Flow
We now consider the effect of a uniform magnetic field that is aligned with the ve
locity field a t the inflow to flow past a cylinder. We have already verified that the
incompressible code obtains the correct answer for the case without magnetic field.
This provides us with a baseline for comparison. The farfield boundary conditions
are straightforward to provide for the magnetic streamfunction. The condition at the
wall was chosen (somewhat arbitrarily) so th a t the magnetic streamfunction is set to
zero there. This condition sets the normal component of the magnetic field to zero
a t the cylinder. A summary of the simulation parameters is given in table 6.4.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
148
mm 1_ t I wlc
Figure 6.4: Incompressible Orszag-Tang vortex ( t= l , instantaneous fields). Top Left: x component of velocity, Top Right: y component of velocity, Middle: Pressure, Bottom Left: x component of magnetic field, Bottom Right: y component of magnetic field.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
149
m m m| r ' r . . X • I !(
ama
<.mm•Ofnesr•last*•L<sm•Lacs•uiora•r7«nj•LMW
[£=1
§H
I aooorroti aaaoi«Mts aooonaw9 .m C E -6S
I 4.WT77&Q3I rsiot&asJ ZMJ23&47 j -Z**SU£M I -4.1C91E40 1.7.| -amiuui I ■aaeoiaaM l-aaoorm u
Lotnxa4.L4M 74& 0* 3 .7.4137IB4V 1 - U M 7 B 4 B f -LfJI77&OB 1 - r s « 7 & o * I • 1 Q M E 4 I I -1717G6&0* I -<JC1A&Oi
Figure 6.5: Incompressible Orszag-Tang vortex ( t= l , instantaneous fields). Top Left: Velocity streamlines, Top Right: Magnetic streamlines, Middle Left: Vorticity, Middle right: Divergence of velocity, Bottom Left: Curl of magnetic field, Bottom Right: Divergence of magnetic field.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
150
Lt*ns
Figure 6.6: Incompressible flow past a cylinder with inflow magnetic fields. From the top: (1) x component of the velocity field, (2) y component of the velocity field, (3) x component of the magnetic field, (4) y component of the magnetic field
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
151The instantaneous primitive variable fields are shown in figures 6.6 and 6.7. We
see from figure 6.7 th a t the shedding pattern at the cylinder is only slightly perturbed
but that the vortex street has been destabilized downstream from the cylinder. The
wake has also become more complicated with more small scale structures appearing
than before.
I a x tm I a n u sI (tIMM
j l«.«R 7 I I4«Z3»
M JIZS
4Miim4ixm•444*447•Last•15*071- lm0.2101>10421•ttJOIl•14.1412■17 4314•I* H U
&i*ou01)147
4 im so ■01X771
Figure 6.7: Incompressible flow past a cylinder with inflow magnetic fields aligned with inflow velocity. Top: Pressure field, Middle: Vorticity, Lower: Stream function of the magnetic field
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 7
Com pressible Flow Sim ulations
The Discontinuous Galerkin method (DGM) that we outline in the next sections is a
variant of the method proposed by Cockburn and Shu et al in a series of papers[68, 69,
70, 67, 71]. Lomtev implemented the method for triangular and tetrahedral spectral
hp elements in [72, 73, 74]. We have extended it to the polymorphic family of spectral
elements. Using the toolkit of element objects we have developed thus far this was
a straightforward procedure because the DGM is a local method. At each stage
of a computation, the only extra data that an element needs to access is from its
neighbouring elements or domain boundary conditions.
We have applied the DGM to the Euler equations, the compressible Navier-Stokes
equations and the Magnetohydrodynamic equations in two and three-dimensions us
ing the array of spectral elements described in chapter 2. In this chapter we will
consider the Euler and compressible Navier-Stokes equations, and in the next chapter
we outline our progress with the compressible MHD equations.
152
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
153We only dem onstrate explicit methods in. this chapter and the next chapter, which
will force A t to be small if the problem is dominated by viscous effects. Oden, Babuska
and Baumann presented a formulation for an implicit solver using DGM in a series of
papers [75, 76, 77, 78, 79]. Their m ethod shows promise if the iterative solver can be
preconditioned well. Their method does not require the use of auxiliary methods but
it does involve a large number of derivatives to be evaluated to compensate for this
saving. It will be interesting to see if the implicit m ethod is cheaper than the explicit
method when the number of iterations per time step axe factored into the total cost.
7.1 Numerical Formulation
For simplicity we first consider an convection equation for a state vector v.
d v dFx dF* dFz n frr xd t + ~ t e + W + ( ]
where F (d) is a function of the state vector.
In the discontinuous Galerkin formulation we consider an approximation space S
which may contain discontinuous functions. The discrete S 5 contains polynomials
within each “element” but zero outside the element. Here the element may be any of
the elements we have described and we denote this element by E{. Thus, the com
putational dom ain Q = (Ji Ei, and i?,-, E j overlap only on edges. Consequently, each
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
154element is treated separately, corresponding to the following variational statement:
Sfio.vlir. dT?x dl?y dFz ~= ~ (w’ (F * - F * K + ( p ’ - F,')» .+ (* “ - F ') n . ) Sa
(7.2)
where w E S 5.
Computations on each element are performed separately, and the connection be
tween elements is a result of the way boundary conditions are applied. Here, boundary
conditions are enforced via the flux F (v ) tha t appears in equation (7.2). Because this
value is computed at the boundary between adjacent elements, it may be computed
from the value of v given at either element. These two possible values are denoted
here as v 1 (internal) and v E (external), and the boundary flux written F^v7, v E).
Upwinding considerations dictate how this flux is computed. For the case of a hyper
bolic system of equations, an approximate Riemann solver would be used to compute
a value of F based on v 1 and v E.
7.1 .1 D iscontinuous G alerk in for D iffusion
We consider as a model problem the parabolic equation with variable coefficient v to
demonstrate the treatment of the viscous contributions:
ut = V • (j/Vu) -F / , in ft, u E L2(Q)
u = <7(x, t), on
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
We then introduce the flux variable155
q = —uV u
and re-write the parabolic equation
ut = —V - q -I- / , in Q.
1 „ . ^—q = —V u, in il v
u = <7(x, t), on d£2,
The variational formulation of the problem is then as follows: Find (q, u) G H (div ; Q) x
L2(Q) such tha t
(u t , w )Ei = (q, Vt«)£i- < w, q B - n >aEi + ( / , w)Ei, Vw G L2{0)
^(q>v) = Vu)Bi- < u B, v n >aEi, Vv G H(grad; fi)
u = g(x.,t), on dQ
where the parentheses denote standard inner product in an element (E{) and the angle
brackets denote boundary terms on each element, with n denoting the unit outwards
normal and:
H(div, f2) = { v G L 2( f i ) | V - w G L 2(fi)}
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
H(grad-,Q) = {u € £ 2(ft)|Vu € L2(fi)}156
The surface terms contain weighted boundary values of v B ,qB, which can be chosen
as the arithmetic mean of values from the two sides of the boundary, i.e. v B = ,
and qB = (qE+qI) .
Integrating by parts once more, we obtain an equivalent formulation which is
easier to implement and it is actually used in the computer code. The new variational
problem is
(ut , w )Ei = ( - V • q, w)Ei- < w , (qB - q1) • n >Ei + ( / , w)Ei, Vm E L 2(Q)
i ( q , v ) = ( -V u , v)Ei— < (u B — u1)n, v >dEi, Vv E H{grad; f2)
u = g (x ,t) , on d£2
7.2 Compressible Navier-Stokes Simulations
The compressible Navier-Stokes equations in conservative form are:
c?S QFEuler dFEuler QYEuler QjpVisc Q~pVisc gj?Visc I s | y | z — x | y j zdt dx dy dz dx dy dz
with following definitions:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
157
Variable Descriptionp(x, t) Density
v (x ,t) = (u, v ,w ) (x,t) VelocityE - j ^ + ^ ( p v v ) Total energy
P Pressure
4 II P3 Tem peratureP r = SEHK Prand tl number
P Dynamic viscosityA Bulk viscosityK Therm al conductivity
p Euler _ px 2 _|_ p? pmu, (E + p)u)T
p Euler _ (pv,f)VU,pv2 + p ,f)V W ,(E + p )v )T
TpEuler _ (pyj: pwU, pwv, /9U/2 + P, (i? + p)w)T
T-.vi5C „ du 3u dv du dwF * = (0’ 2^ + A V - v ' ^ + ^ ’^ + f , & -
p d v 2 l d T . T A ( V . v ) u + M ( v - V ) « + ¥ — + - ^ )
dv du „ dv dv dwF » “ (0- ^ + ^ v 2^ + A V ' v ’f‘& + f ‘ v
A ( V . v ) „ + (1( v . V ) . + f | - + i | ) r
du/ du dw dv n dw , _F - = (0' ^ + ^ ^ - + ^ > 2' i 9 7 + AV ' v ’
A(V • V)w + „ ( v • V)w + f
The formulation of the Riemann solver for the Euler fluxes is identical to tha t
proposed by Lomtev [80]. We will omit them here for brevity. We will go into more
detail for the compressible MHD equations tha t we develop in the next chapter. The
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
implementation of the upwinding routine is independent of the element type.158
7.2.1 C onvergence
We first consider the convergence rate of this new formulation by solving for an
inviscid and isentropic flow problem in the geometry shown in figure 7.1. Low-order
methods erroneously produce entropy from inlet to outlet for this problem. Here we
show in figure 7.1 (bottom) that the entropy error converges exponentially fast to
zero with p-refinement. A comparison is shown on the right plot between a fully
unstructured and a hybrid discretization; more elements are used in the unstructured
grid.
We repeated this test with the three-dimensional compressible code. In figure
7.2 we show the three-dimensional domain that we extruded from a two-dimensional
bump mesh with K = 120. Large hexahedra are used at the inlet and outlet, whereas
smaller prisms and hexahedra are used around the bump. We also show th a t the
entropy again decreased exponentially fast with increasing expansion order.
7.2.2 C ylinder F low
For this test of the hybrid incompressible code we consider two-dimensional flow past
a circular cylinder. The cylinder has unit diameter and the domain surrounding the
cylinder is a rectangle [—22,69] x [—22,22]. Uniform upwind boundary conditions
are employed at the box. No slip and wall tem perature are imposed on the cylinder.
Figure 5.5 shows the domain, mesh and boundary conditions. This is the same mesh
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
159
2
1.5
>» 1
0.5
0
0.5
10 '* r
1°'* r
r rcUl~104 - J
lff* r
io-' r
Figure 7.1: Density contours obtained on a hybrid grid for an inviscid M = 0.3 flow (Top), on a triangle grid (Middle). The bottom plot shows exponential convergence of the error for the unstructured (triangles) and the hybrid (squares) grid.
5 10Expansion order
X
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
160
Expansion Order
Figure 7.2: Inviscid M = 0.3 flow past a bump in a three-dimensional (periodic in the spanwise direction) domain. From the top: (1) domain, (2) spectral element mesh used in the convergence test K = 120, (3) Iso-Contours of density and (4) convergence of entropy to zero with increasing order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
161
Param eter ValueDimension 2d
Re 100, 150, 200 and 250Mach 0.5 and 0.7
A t 0.05 to 0.0002N-Range 8 to 10
K th 332KQuad 448
Method Discontinuous Galerkin
Table 7.1: Simulation param eters for compressible flow past a cylinder.
Strouhal FrequencyRe Incompressible N s k 'T ocr Compressible M skTTocr (Mach 0.5)100 0.1662 0.1611150 0.1853 0.1812200 0.1972 0.1934250 0.2054 0.2020
Table 7.2: Comparison of Strouhal frequency of incompressible against compressible (Mach=0.5) flow past a cylinder.
that was used for the incompressible cylinder flow tests.
The mesh has 332 quadrilaterals and 448 triangles. It was designed to direct
resolution around the cylinder, have regular resolution in the wake behind the cylin
der, and block out with large elements to push the farfield boundaries away from
the cylinder. We ran the simulation at N = 8 for Re = 100,150, 200 and 250. A
summary of the simulation param eters is given in table 7.1. In table 7.2 we show that
the Strouhal number is only changed by about three percent by the compressibility
at Mach 0.5.
We also ran the cimulation a t Mach 0.7 in order to compare with the results of
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
162
Code Method St Q Outflow (-§)dgL [73] Discontinuous Galerkin 0.158 1.846 0.246 35
//flow [81] Explicit Galerkin 0.159 1.843 0.248 35AIe kT ocr Discontinuous Galerkin 0.157 1.802 0.255 69
Table 7.3: Comparison of Strouhal frequency, drag and lift coefficients from hybrid A /e k T ocr with results from other methods. Re=100, M=0.7
Lomtev [73] who used Discontinuous Galerkin with spectral element triangles and
Beskok [81] who used Galerkin on quadrilaterals. We show good agreement in table
7.3 for the Strouhal frequency, however the drag (Cd) and lift (Ci) coefficients are
different by about three percent. This difference can be accounted for by the different
dimensions of the meshes used. For the A /e k 'T'ocr run the outflow was 69 diameters
from the cylinder, whereas for the other methods the outflow was only 35 diameters
from the cylinder. When we ran with the smaller mesh it was obvious tha t there
was a higher level of noise induced when the von Karman street exited the domain.
The comparision with Lomtev shows that using quadrilaterals has not adversely af
fected the polymorphic implementation of the Discontinuous Galerkin method. The
comparison with Beskok shows that the Discontinuous Galerkin method on unstruc
tured polymorphic grids compares well with the more traditional Galerkin methods
on semi-unstructured quadrilateral meshes using collocation.
In figure 7.3 we show instantaneous iso-contours for compressible flow past a
cylinder at Mach=0.5. Also in figure 7.4 we show vorticity which is quite similar to
the incompressible case shown in figure 5.6.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
163
Figure 7.3: Instantaneous iso-contours for the simulation of compressible flow past a cylinder {Re = 100, M = 0.5) From the top: (1) density, (2) pressure field, (3) x component of the velocity field, (4) y component of the velocity field
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
164
8
0
•s
•to
*
Figure 7.4: Instantaneous iso-contours of vorticity for the simulation of compressible flow past a cylinder {Re = 100, M = 0.5)
7.2 .3 F low P a st a T w o-D im en sion a l N A C A 0012 A irfoil
We simulated compressible flow at Mach=0.5, Re=10,000 (based on chord length)
past a NACA0012 airfoil a t zero angle of attack. This airfoil needs to have good res
olution at the leading and trailing edges and to have good boundary layer resolution.
We used thin quadrilaterals on the body, and small elements a t the nose and tail then
blocked out using triangles in the fax domain and a regular array of quadrilaterals in
the wake. It is im portant to keep the resolution smoothly varying in the wake since
abrupt changes in the resolution there can lead to noise generation.
4180 quadrilaterals and 3307 triangles were used in the mesh which was created
by Kirby [82] using SIMPLEX2D [83] and an advancing front/bioeking algorithm. A
summary of the simulation parameters is given in table 7.4.
In figures 7.6, 7.7 and 7.8 we show instantaneous iso-contours of density, divergence
of momentum, and curl of momentum, respectively, in the near wake of the airfoil.
The divergence of momentum is equal to minus the rate of change of density due to
the conservation of mass equation. Hence, we should expect to see noise generated
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
165
Parameter ValueDimension 2d
Re 10,000 based on total chord lengthMach 0.5
A t 0.001 to 0.00001N-Range 1 to 11
K th 3307K qucuI 4180
Method Discontinuous Galerkin
Table 7.4: Simulation param eters for compressible flow past a NACA0012.
by the m ethod appearing in this field. The plot shows th a t this calculated quantity
is smooth showing tha t at this resolution we are well resolved by this measure. The
vorticity plot, however, shows some noise near the vortices in the wake, so in this
case the vorticity is probably a good measure of accuracy and a small increase in the
resolution should reduce this noise to acceptable levels.
7.2.4 F low P a st a M u lti-B o d y A irfoil
We simulated compressible flow a t Mach=0.5, Re=10,000 (based on total chord
length) past a multi-body airfoil a t zero angle of attack. The small scale features
(i.e. nose and tails of each component) need to be resolved as well as the wake region
and boundary layers. We used thin quadrilaterals on the components, and small ele
ments at the noses and tails then blocked out using triangles in the far domain and
a regular array of quadrilaterals in the wake.
1678 quadrilaterals and 3699 triangles were used in the mesh which was created
by Kirby [82] using SIMPLEX2D [83] and his own advancing front/blocking routines.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
166
10
5
0
•5
-1030-10 0 10 20
X
Figure 7.5: Top: Mesh of full domain for simulation of compressible, Mach 0.5, Re=10,000 flow past a NACA 0012 airfoil at zero angle of attack to the mainstream flow, Middle: Mesh around body and wake, Bottom Left: Close up of the airfoil, Bottom Right: Close up of part of the wake region.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
167
Figure 7.6: The wake region (from f = 2 to ^ = 5) of a NACA 0012 airfoil at zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000, 3307 triangles, 4180 quadrilaterals, N = l l . Instantaneous iso-contours of the density are shown.
Figure 7.7: The wake region (from ^ = 2 to f = 5) of a NACA 0012 airfoil at zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000, 3307 triangles, 4180 quadrilaterals, N = l l . Instantaneous iso-contours of the divergence of momentum are shown.
Figure 7.8: The wake region (from ^ = 2 to ^ = 5) of a NACA 0012 airfoil at zero angle of attack to the mainstream flow. Mach 0.5, Re = 10,000, 3307 triangles, 4180 quadrilaterals, N = l l . Instantaneous iso-contours of the curl of momentum are shown.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
168
Param eter ValueDimension 2d
Re approx. 10,000 based on to ta l lengthMach 0.5
A t 3e-5N-Range 4
Krri 1678KQuad 3699
Method Discontinuous Galerkin
Table 7.5: Simulation parameters for compressible flow past a m ultibody airfoil.
A summary of the simulation parameters is given in table 7.5.
The final two-dimensional example is a simulation of flow past a multi-element
airfoil shown in figure 7.9. The hybrid grid consists of 1638 quadrilaterals and 3739
triangles. The grid was constructed in three phases. Quadrilaterals were created using
an advancing front method from each airfoil. A block of structured quadrilaterals was
used to cover the wake region. The rest of the domain was covered with unstructured
triangles using SIMPLEX2D. This is a Delauney-based grid generator with Steiner
triangulation for guaranteed quality triangulation [4]. The Reynolds number is Re =
10,000 based on the to tal length of the multi-element airfoil, and the freestream
Mach number is M — 0.5. A third-order expansion was used uniformly in each
element everywhere in this problem. In figure 7.10 we plot instantaneous iso-Mach
contours and streamlines around the airfoils. The flow is unsteady as it is evident
from the vortex street present in the near-wake.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
169
0 0 .0 5 0.1 0 .1 5 02x
Figure 7.9: Hybrid Mesh and close ups for the simulation of compressible flow past a multi-body wing
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
170
0.65846505544970.4505290.3465610.2425930.1386250.0346565
0.2
0
- 0.2
-0 .40 0 . 5 1
x
Figure 7.10: IsoMach. contours and streamlines for M = 0.5 flow past a two- dimensional multi-component wing.
0.658465
0.4505290.3465610.2425930.1386250.0346565
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
171
Parameter ValueDimension 3d
Re 2000 base on chord lengthMach 0.5
A t le-4N-Range 1 to 4K P rism s 1960
K Hex 2095Method Discontinuous Galerkin
Table 7.6: Simulation parameters for compressible flow past a NACA0012 airfoil with endplates.
7.2.5 F low P a st A N A C A 0012 W in g W ith E nd p lates
The first three-dimensional problem we consider is flow past a NACA 0012 airfoil
with plates attached to each end. We impose uniform upwind boundary conditions
a t inflow and outflow, and the domain is periodic from one end of the airfoil to the
other.
This is a simple model of a wing between an engine and fusalage. The domain was
meshed with SIMPLEX2D [83] and extruded using 1960 prisms and 2095 hexahedra.
A thin boundary layer of hexahedra was used on the surface of the wing and the
simulation was run with upto 4</l order expansion and Re=2000 (based on chord
length). A summary of the simulation parameters is given in 7.6.
At low-order the simulation ran to steady state. This is due to a reduced effective
Reynolds number achieved because of numerical dissipation. As we increased the
order we saw unsteadiness developing in the wake of the wing, and what appears to
be oblique shedding. This is only a marginally three-dimensional domain but it does
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
172
demonstrate the ability of f s f stz'T'c l t to direct resolution into boundary layers and
to fill out a domain w ith larger elements.
Figure 7.11: Skeleton mesh for flow past a three-dimensional NACA0012 airfoil with endplates.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
173
Figure 7.12: Iso-contours for x-component of momentum for M = 0.5 flow past a three-dimensional NACA0012 airfoil with endplates.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
C hapter 8
Com pressible V iscous
M agnetohydrodynam ics
Sim ulations
We have already described an incompressible plasma in chapter 6. In this chapter
we generalise the equations to describe a compressible plasma. We implemented the
algorithms in the J ^ b k 'T olt code so that simulations can be run on a broad range
of parallel computers using the Message Passing Interface (MPI) [60]. We tested the
code on some simple magnetohydrostatic test cases and on the Orszag-Tang vortex.
Also, we sim ulated flow past a cylinder with an imposed far-field magnetic field.
174
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1758.1 Formulation
The equations for compressible magnetohydrodynamics can be expressed in conser
vative form in compact notation as:
- V • (pv)
- v • ( p w ‘ - B B ' + (p + i |B |2) / - -~ r )
—V x (B x v 4- -^-V x B)Or
- V • ((£ + p )v + ( i |B |2I - B B ‘) . v - i y - r
+ j -(B ■ VB - V ( | lB |2)) - ~ ^ T )
0
(djVi + diVj) - | v • vSi:i
Alternatively, in flux form with the explicitly stated fluxes as:
dXJ dF Euier dF Euler dFEuler dJ?Visc dFVisc Q~pVisc~dt = dx dy dz + dx + dy + dz
V -B = 0
U = (p ,p u ,p v ,p w ,B x,B y,B z, E )
FEuler = pu2 - B l + p, puv - B xB y, puw - B XB Z, 0, uBy - vB x, u B z - wBx,
(E + p )u — (v • B )BX)T
FEuler = (pv, pvu - B yB x, pv2 - B 2 + p, pvw — B yB z, vBx - uB y, 0, v B z - wBy,
dp_dt
d(frv)dtdBdtd Edt
V B =
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
(E + p)v — (v • B )By)T176
■pEuier _ /JWU _ b zB x, pyjy _ B zB y, pw2 — B 2 + p, wBx — uB z, w B y — vB z, 0,
(E + p)w — (v ■ B)BZ)T
f 1'*” = <0,— <— - - v - V) - i — + ? y - ) — (— + — )x ’ S v dx 3 ’ Sv dy dx S v dz dx
0 L (^ y _ dB-) 1 (dB * dB ^’ 5 / dx d y U Sr Kdx d z J’1 / 2 , _ , „ l d v 2 1 dT. I T d f l B l 2) „ . , T
^ ( - 3 (V - v)u -F v • Vu + 4- - — ) + - ( _ ---------B - V B,))
F w,c _ ,n _i_f 9v_ du 2. r f t L _ I v 7 f 1 r d v , d w \y S v ^dx + d y ^ S ^ d y 3 S v d z dy
1 d B x dBy 1 dBz dBySr dy dx ’ 5 / dy d z h1 f 2 r ^ t~7 l ^ v 2 1 d T , 1 ,1 0 ( |B |2) „5 ( - 5 (V • v ) . + v ■ V , + - - g - + - w ) + - ( — B • V B ,) )
revise rn 1 ,dw du 1 dw dv 2 dw 1p - = (0’ i ; {^ + 5 I ) ' ^ ( ai7 + a ; )' ^ (^ - 3 V v ) '
1 d B x dB z I dBy dB z Sr K d z d x h STKdz d y }’ '1 / 2 ,t7 vr 1 d v 2 1 dT . l , l d ( | B ] 2) „^ ( - 3 (V • v )« + v - Vw + - w + j - r ^ ) + B • V S .))
with the variables and parameters defined in table 8.1.
8 .2 The V • B = 0 Constraint
The presence of the V -B = 0 constraint implies th a t the equations have a semi-elliptic
character. It has been shown in [65] that even a small divergence in the magnetic
fields can dramatically change the character of results from numerical simulations.
An alternative approach to the magnetic stream function which we used for the
incompressible MHD simulations was developed in [62]. This modified the equations
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
177
Variable Descriptionp(x, t ) density
v(x, t) = (u,v,w)(-x,t) velocityB(x, t) = (B x , B y, Bz)(x, t) magnetic fields
E — + 2 Cpv - v + B - B) total energyP Pressure
P = P + • B Pressure plus magnetic pressureT ----E_Rp Temperature
P r = ^K Prandtl NumberR Ideal gas constantV Magnetic resistivityP Viscosity
C _ Po vA L.0v p. Viscous Lundquist numberSr = ^ Resistive Lundquist number
% Specific heat at constant pressure
CS II •S Alfven wave speed
ii Alfven Number
Table 8.1: Variables and parameters used in the equations of compressible MHD.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
178by adding a source term proportional to V • B:
Spoweii = - ( V • B )(0 ,B x,B y,B z, u , v , w , v - B ) T
to the right hand side of the evolution equation. In [62] this term was incorporated
into the Riemann solver for the Euler flux terms. We evaluate this term as a source
term, without modifying the Riemann solver. The divergence of the magnetic field is
calculated using the DGM derivatives of the magnetic fields.
8.3 Implementation of the Euler Flux Terms
We now apply the same techniques tha t we used for the Navier-Stokes equations.
We evaluate these fluxes and their derivatives in the interior of the elements and add
correction terms for the discontinuities in the flux between any two adjacent elements.
In order to evaluate the Euler flux a t an element interface we use a one-dimensional
Riemann solver to supply an upwinded flux there. At a domain boundary we expect
the user to provide far field conditions and treat the exterior boundary as the bound
ary of a “ghost” element. This way we can use the same Riemann solver at all element
boundaries.
We linearize the one-dimensional flux F f “/er in the normal direction to a shared
element boundary using the average of the state vector at either side of the element
boundary. That is, since F&uler is a nonlinear function of the state vector we use the
average state to form an approximation to the Jacobian of the flux vector A c.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
179The Jacobian m atrix for the flux vector for the evolution equations expressed in
primitive variables is simpler than the conserved form. Thus, we will perform the
linearization about the primitive form and transform to the conserved form. The
primitive Jacobian m atrix A p has the form:
u p 0 0 0 0 0 0
0 u 0 0 Br Ex B~ 1
P p P p
0 0 u 0 - E x B r 0P P p
0 0 0 u B zP
0 B zP
0
0 0 0 0 0 0 0 0
0 By ~ B X 0 — V u 0 0
0 B z 0 ~ B X —w 0 u 0
0 IP 0 0 — (7 — l)u • B 0 0 u
The left and right eigenvectors of the primitive Jacobian m atrix A p, due to Powell
et al [63], are:
Entropy wave:
Ae = u
le = ( 1 ,0 ,0 ,0 ,0 ,0 ,0 , - 4 )a1
r e = (1, 0, 0, 0, 0, 0, 0, 0)‘
Alfven Waves:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Fast waves:
A f = u ± c /
l, = 5ij(0,±a/c/,Ta,c,/3I ,T«,c.A,i8„0,2^5,2^2,2i)2a y/P y/P P
Tf = ( p a f , ± a f C f , ^ f a scsp xP y , T a scsPxPz, 0 , a sPya y / p , a sPzay/p,af 'yp) t
Slow waves:
As = u ± cs
l. = ± ( 0 , ± a , c , , ± a / cf /3x/3s, ± a / cf & & A - ? ^ , - ? , - )Z a V P V P P
rs = (pots, ± a scs, ± a f cf pxPy, ± a f cf /3x/3z, 0, - a f j3yaVp, ~<^fPzaVp, Ois'yp)t
Where:
_ 7p -F B • B(a*)2
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
181
a 2 a 2
7 C /_ c ?2 ^ ~ g2
Px = sgn(Bx)
PyBy
Z
We can transform between the primitive and conserved variables with the following
transform:
<9U „ d WA c — d W p d V
where:
U = (p, pu , pv, pw, B x, By, B z, E)
are the conserved variables and:
W = (p, u, v, w, B x, By, B z,p)
are the primitive variables. This gives:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
182
d U
d W
1 0 0 0 0 0 0 0
u p 0 0 0 0 0 0
V 0 P 0 0 0 0 0
w 0 0 p 0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
v - v2 p u p v p w B x B y B z 1
7-1 .
and
a wa u
1 0 0 0 0 0 0 0
_XL
p
1p
0 0 0 0 0 0
_y_
P0 I
p0 0 0 0 0
_w
p0 0 I,
p0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
f v - v —7 u — 7 u — ■yw - 7 B x -7 B y ~ l B z 7
where 7 = 7 — 1.
We are now in a position to evaluate the upwinded flux at the element boundaries.
We use following formulation for the upwinded flux:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
183
1 9 U k=?F(Uj, U s) = ~ (F (U r) + F ( U £ ) - —
d Wo* = Ik * - ^ j ( U s - U r)
Here the lk and r k are the ordered left and right eigenvectors of the primitive Jacobian
matrix. We have to apply the operator to the right eigenvectors to calculate the
conserved flux. The A* are the wave speeds associated with the eigenvectors.
8.4 Implementation of the Viscous Terms
The viscous terms are evaluated in two steps. First, we obtain the spatial derivatives
of the primitive variables using the discontinuous Galerkin approach. Then, we repeat
the process for each of the viscous fluxes using these derivatives. If the user supplies
Dirichlet boundary conditions for the momentum and energy variables we set these
terms explicitly after the fluxes have been evaluated and then project the result using
the orthogonal basis.
8.5 Summary of the Compressible MHD Code Im
plementation
A possible implementation of the algorithm is:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
184• Step 1: Read in initial conditions U(x, 0), evaluate the fields at the element
quadrature points. Set n = 0.
• Step 2: Calculate the upwinded fluxes Fn at the Gauss quadrature points on
the element interfaces. At domain boundaries use the prescribed boundary
conditions for the exterior values of the fields.
• Step 3: Calculate the Euler flux terms: F ^uier, F ^ “ier and F f uler at the element
quadrature points.
• Step 4: For each component of the state vector Uk = (U(x, tn))k calculate
.Q - p E u U r g p E u / e r g p E u l e r
1 d x + d y ”*■ d z >k ■
• Step 5: Interpolate the upwinded flux to the correct quadrature for the ele
ment interface. Scale the fluxes with the edge Jacobian divided by the volume
Jacobians. Add this to the divergence of the Euler fluxes calculated in Step 4.
• Step 6: Take inner product of the fluxes with the orthogonal basis. Evaluate
the resulting polynomials at the quadrature points.
• Step 7: Calculate the spatial derivatives of the primitive fields. For example,
we will need We evaluate this as described in section 7.1.1 using the DGM
to compensate for the derivative across element interfaces, and the average of
u at the element interface.
• Step 8: Use the derivatives of the primitive fields to construct the Viscous flux
terms: F^'sc, F^'5C and F ^ isc.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
185• Step 9: Take the divergence of the viscous flux terms.
• Step 10: Subtract the divergence of the Euler fluxes obtained in Step 6 from
the divergence of viscous fluxes obtained in Step 9.
• Step 11: Take inner product of the result from Step 10 with the orthogonal
basis. Evaluate the resulting polynomials at the quadrature points and place it
in Uf(x, £n" 9)
• Step 12: Update the state vector U(x, £n+1) = U(x, tn) + A t 5D?/39Uf(x, tn~q)
using an Adams-Bashforth integration scheme.
• Step 13: Increase n by one. If tn is less than the term ination time return to
Step 2.
• Step 14: O utput final values of the state vector U(x, tend) .
8.6 Two-Dimensional Magnetohydrostatic Exam
ple
A simple test for the compressible MHD component of the M s k T ctr code is to
consider a steady irrotational magnetic field and zero velocity. The test was performed
as an initial value problem and the following exact solution:
p = 1
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
186u = 0
v — 0
e(-2*ry)£ = 19.84 + ---------
2
Bx = — cos(‘irx)e(- ~ ^
By = sin(7ra;)e^-iry^
was used as the boundary conditions and as the initial condition. This solution was
derived by Priest [84]. The irrotational magnetic field implies tha t the Lorentz force
is zero so the momentum equations are trivially satisfied. The magnetoviscous term
is zero and the v x B term is also zero. Thus, the compressible MHD equations are
satisfied.
The domain and discretization we used is depicted in figure 8.1. We also show
that the approximation decreases exponentially with increasing expansion order.
8.7 Three-Dimensional Magnetohydrostatic Exam
ple
We modified the two-dimensional test case, used in the previous section, to be a
three-dimensional test case for the compressible MHD component of the N ek'Totr
code. Again we used a steady irrotational magnetic field and zero velocity solution
to the MHD equations. The test was performed as an initial value problem and the
following exact solution:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
187
0 .7
0.6
0 .4
O.t
X
0.8
0 .7
0.6
0 .4
0.1
X
10"
10"
O m f \-7 fc10w
10 s
10 '
1 1 1 ' ' 1 ' 1 1 ' ' 1 1 ' ' ' ' 1 ' ' ' ' I ' '4 6 8 10 12
Expansion Order
Figure 8.1: Magnetohydrostatic test case for the compressible code. Top Left: Mesh {Kth = 38,KQuad = 22) Top Right: Magnetic streamlines of steady solution at N = 12, Bottom: dependence of steady state error on expansion order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
188
1
0
0
0
sin(ax)(e~az — e_ay)
—sin(ay)e~ax — cos(ax)e~ay
(cos(ax) — cos(ay))e~az
5 + 0 .5(B | + f l J + B j )
27rT
was used as the boundary conditions and as the initial condition. By construction,
the magnetic field is irrotational and the Lorentz force is zero so the momentum
equations are trivially satisfied. The magnetoviscous term is zero and the v x B term
is also zero.
The domain and discretization we used is depicted in figure 8.2. We also show
that the approximation decreases exponentially with increasing expansion order.
P =
u =
v =
w =
B x =
By =
B z =
E =
a =
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
189
: a s x j m0357647 0307315 00*64137
4X113345 -0173576
i -0434006 •0544337 •0734661 -0314999 -L07333
Expansion Order
Figure 8.2: Three-dimensional magnetohydrostatic test case for the compressible code. Iso-contours of the magnetic and energy fields axe shown at time t = 1. Top Left: Tetrahedral mesh used, Top Right: x-component of the magnetic field, Middle Left: y-component of the magnetic field, Middle Right: z-component of the magnetic field, Bottom: Convergence plot showing exponential decrease in L 2 error with increasing expansion order.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
1908.8 Compressible Orszag-Tang Vortex
The Orszag-Tang vortex is an initial value problem in a squaxe periodic domain
(length L). It has been investigated by [66] and [64] amongst others . It demonstrates
that turbulent scales can result from a coherent initial condition with only two spatial
frequencies.
The initial conditions we used were:
p = 1
u II 1 Co <s>,
to «ci
V = sm { L )
B x = sm{ L )
By. .47TX.
= sm { )
p ^ 1 A= C + -cos(- 4
S-k x . 4 Arrx. 2iry. .27ra;. .27ry. 1 , 47n/.-) + -c o s (-^ -)c o s (— ) - cos(— )cos(— ) + -cos{— )
where C fixes the initial background mach speed and p is the instantaneous pres
sure for the equivalent incompressible flow. A summary of the simulation parameters
is given in table 8.2, and figure 8.3 shows the mesh used for this case.
In figure 8.4 we show iso-contours of the conserved variables at t = 1. We see tha t
using an expansion order of nineteen shows very smooth and symmetric fields.
We now consider the effect of expansion order and using an unstructured grid on
accuracy. Vorticity is a good indicator of noise in the solution of unsteady simulations.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
191
6
5 ts:isi
4
tic198!issi3
2
1
00 2 53 4 6X
Figure 8.3: Mesh used for the compressible Orszag-Tang vortex simulations on a structured mesh.
Parameter ValueDimension 2d
5 , 100Sr 100
A (Alfven Number) 1.Mach 0.5
N 19KQuad 64
Method Discontinuous Galerkin
Table 8.2: Simulation parameters for the compressible Orszag-Tang Vortex.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
192
14.101713.650811212.7491121982111474IL396510.945710.494810.0449193119.142268.69148.24054
L 27227114041108531.17665L 144781.1129LOB 1(231.049161.017280.9854080.9535340.921660.8897860857912
(.06259
•0.104175
■0.437537•0.604218
• 1.10426
'II, W: ..HUis: iJM A ' S . i .
1.414751.192830.9709060.7489850J270630.3051420.0832205
■0.1387014X3606224X5825444X804465•1.02639
114831•1.47023
'&r mmm nw .•' - <: ■'P l i - /d i r w
MUrr1 P>l -
1.414411.192540.9706710.748803
•0.138667
•1.02614•114801•1.46987
(.1414
i iS S R0.783313
014618400671411
•0.111902
•0.469988•0.649031•0.823073
.00712• I .18616
Figure 8.4: Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.5). Top left: density, Top right: energy, Middle left: x-component of momentum, Middle right: y-component of momentum, Bottom left: x component of magnetic field, Bottom right: y component of magnetic field.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
193
OaXr)M0615
0.7130160.139732
L3S0122.13341
>172669-319997•317326
02434310.1926420139030.0870244
-00714026•0124212-0177021
-
6.97376347611
3.076811.777160477314
-112179-3.42144-4.72109
-7J2039
-9.91969
■ i ■ ■ i ■ ' ■ *■
Dhrffl)I 0.00404916 I 000344398 j 00028313
000223362 | OOOI62U4
000102326 j 000041108
3 >000011707 1-000079227 I -0.00139746 R -0.00200264 1-0.00260782 I >0.003213 I -0.00381118
>»
X
Figure 8.5: Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.5). Top left: curl of momentum, Top right: divergence of momentum, Middle left: curl of magnetic field, Middle right: divergence of magnetic field, Bottom left: momentum streamlines, Bottom right: magnetic field streamlines.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
194Low resolution will show up as non-smoothness in vorticity since it is the derivative
of the computed fields. In figure 8.7 we compare the vorticity at time t = 1 for a 132
triangle mesh (shown in figure 8.6) run w ith N = 4 and N = 6. The top figures show
how the vorticity varies across the lower-left to top-right diagonal. The vorticity in
this direction should be symmetric about the mid-point. We see that a t N = 4 the
profile is noisy, the peeks are not well resolved and the symmetry is not very well
represented. At N = 6 the symmetry is better but the profile is still quite noisy. In
the next figure 8.8 we see at N = 10 th a t the symmetry is very good but the slopes
are still a little rough. Finally, at N = 16 we see tha t the sym m etry is perfect and the
profile is very smooth. We also see that the smoothness of the iso-contours improve
with each increase in expansion order.
6
5
4
3
2
1
oo 1 2 4 5 63X
Figure 8.6: Compressible Orszag-Tang vortex triangle mesh with K=132.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
195
x x
>%
X X
Figure 8.7: Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.2, K=132). Top Left: Curl of momentum along the diagonal(N=4), Botton Left: Iso- contours of curl of momentum(N=4), Top Right: Curl of momentum along the diag- onal(N=6), Bottom Right: Iso-contours of curl of momentum(N=6)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
196
x x
x
>•
X
Figure 8.8: Compressible Orszag-Tang Vortex ( t= l , instantaneous fields, Mach=0.2, K=132). Top Left: Curl of momentum along the diagonal(N=10), Botton Left: Isocontours of curl of momentum(N=10), Top Right: Curl of momentum along the diagonal(N=16), Bottom Right: Iso-contours of curl of momentum(N=16)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
197
Parameter ValueDimension 2d
s v 100Sr 100A 0.1
Mach 0.5A t le-4
N-Range 1 to 8K Tri 490
Method Discontinuous Galerkin
Table 8.3: Simulation parameters for compressible flow past a cylinder with a magnetic field.
8.9 Flow Past a Cylinder
As with the other applications we have developed using the spectral element toolkit
we resort to cylinder flow as a test case. We see that the fields in figures 8.9 and 8.10
closely resemble the equivalent fields in the incompressible version of this simulation in
section 6.5. This is no surprise since we chose a fairly low Mach number (Mach=0.5)
to run this test a t. We can also compare this to the compressible case without a
magnetic field 7.2.2. We see th a t up-down symmetry of the x-component of velocity
has been broken in the wake and the regular patterm of the y-component of velocity
also breaks down ten diameters from the cylinder. This is clear evidence tha t the
magnetic fields are indeed causing the von Karmann street to become unstable with
time.
A summary of the simulation param eters is given in 8.3.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
198
to
Figure 8.9: Instantaneous iso-contours of the simulation fields for flow past a cylinder (Sv = 100, Sr = 100, A = 0.1) with a magnetic field. From the top: (1) density, (2) pressure, (3) x component of the velocity field, (4) y component of the velocity field
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
199
Figure 8.10: Instantaneous iso-contours of the simulation fields for flow past a cylinder (St, = 100, Sr = 100, A = 0.1) with a magnetic field. Top: x component of the magnetic field, Bottom: y component of the magnetic field
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Chapter 9
Sum m ary
In this thesis we have achieved the following goals:
• Developed a natural approach to implementing polymorphic elements ( hp finite
mixed-elements) using an object oriented language.
• Developed efficient algorithms to align the coordinate systems of polymorphic
elements. (Reducing NP hard problems to linear algorthms)
• Quantified theoretically and numerically the approximation properties of the
orthogonal bases on these elements.
• Applied polymorphic element methods to the following physical problems:
— Incompressible Navier-Stokes in two- and three-dimensional domains.
— Incompressible Navier-Stokes in two- and three-dimensional moving do
mains.
— Incompressible Magnetohydrodynamics in two-dimensional domains.200
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
201— Compressible Euler in two- and three-dimensional domains.
— Compressible Navier-Stokes in two- and three-dimensional domains.
— Compressible Magnetohydrodynamics in two- and three-dimensional do
mains.
The Navier-Stokes solvers are direct extensions of the work of Sherwin [30] for
the incompressible code and Lomtev [72, 73, 74] for the compressible code. The new
development here is to allow all the polymorphic element types to be used as well as
the triangles and tetrahedra reported there. This is a significant advance th a t will
facilitate the move to even higher resolution because thin boundary layers can now
be resolved. This advance did not come in terms of a loss in efficiency. We applied
object oriented programming methodology to make the management of varied element
types as transparent as possible, with no perfomance hit. Also, since the companies
tha t make compilers are now concentrating on optimizing C-H- code, Afetz'Tocr will
benefit from their new developments.
We think that this is the first time high-order polymorphic elements have been
applied to the equations of conservation laws, and especially the MHD system. The
equations are quite complex and without the natural programming approach we de
veloped, it would have been difficult to construct a robust code for them.
Now that N e k 'T clt has been developed to be a powerful tool the question re
mains as to what kind of physical problems can be simulated with it. We have shown
it capable of handling simulations with millions of degrees of freedom when run on
parallel super-computers. The discontinuous Galerkin modules only utilize pairwise
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
202communication, i.e. no piece of data needs to be communicated between more than
two processors. This form of communication can be efficiently implemented and good
scaling can be achived, i.e. the overhead per extra processor does not increase rapidly
as the number of processors is increased. This means th a t a discontinuous Galerkin
simulation can be spread over many processors efficiently and therefore A fe n 'T olt
should be capable of solving simulations with several orders more degrees of freedom
as processor power increases and many-processor machines are built. We predict that
two- and three-dimensional simulations of flow past complex two-dimensional bodies
at up to Re = 106 should be possible a t the turn of the millenium.
For three-dimensional simulations the potential to reach these goals is there, how
ever some advances will have to be made to ease the data jam that results from the
huge numbers of data points created. Now that the simulation code A fe n 'T a r has
been written, simulations can be started once the domain has been meshed. However,
the data mining of results will be a significant challenge. In future work the use of in
telligent software agents for extracting relevant data on the fly should be investigated.
These same agents will also be necessary to direct the available resolution towards
interesting parts of the flows. This data migration in itself will need new approaches
to balancing the simulation over multiple processors. This task will be helped by
the OOP nature of A fe tzTo tr , since each element can be enabled to replicate itself,
delete itself or move itself between processors through a simple interface.
For example, an agent could “tell” an element it is no longer needed. Then it will
communicate to its neighbours th a t it is going out of existence, transmit any data
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
203that will be needed to them and then delete itself. This would have been quite a
difficult task with plain C code, and even harder with Fortran code but C + + should
make this a m atter of planning the design of the agent wisely.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A ppendix A
C ontinuous Expansion B asis
A .l Jacobi Polynomials
(1 - * ) ( ! + x) + [(/? - a) - (2 + a + 0)x] dy{x)dx= -A ny(x)
or
d_dx
(1 — ar)1+tt(l 4- x )1+&dy(x)dx
= -A n( l - x ) Q(l + x)py(x)
An = n (n -+- a + j3 + 1)
y ( x ) = Pn P(x )
In the following appendix we document the complete polynomial expansions for
both the m odal and nodal basis discussed in chapter 3. The bases are defined in terms
of three orders P i, P2, P3 corresponding to the local coordinate system. For the modal
expansions we have indicated the indices over which the bases should be assembled
204
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
205in order to achieve the maximum polynomial space although it is possible to reduce
the interior, face and edge expansion orders to produce a variable order expansion as
discussed. The nodal basis must be assembled over all the indicated indices for the
expansion to spam a complete polynomial space.
A.2 Modal Basis
In the following section refers to the ith Jacobi polynomial defined in section
A.
A .2.1 T w o-D im ension al E xpansions
Quadrilaterail D om ain
sJ L
c D
0 r
A B
Figure A .l: Vertex labelling for the standard quadrilateral region
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Using the labelling given in figure A .l the quadrilateral basis is:206
Vertex A : (1=») (1=5)
Vertex B : (l± s) (1=5)
Vertex C : ( » ? ) ( “ )
Vertex D : (l± s) (1±4) ,
Edge AB : (1=4) (l±e) fW (a) (1=4) (0 < t < P i)
Edge CD : ( l= s ) (1±») p W (o) (1±4) (0 < i < P ,)
Edge AC : (1=4) (1=5) (l±it) pM l(6) (0 < j < P2)
Edge BD : (1±4) (1=5) (I±5) pM (6) (0 < j < P 2),
Interior : (l=*) (l± s) P t\ (a ) (1=5) (1±5) P ^ ( b )
(0 < i , j ; i < P i , j < P2)-
Triangular D om ain
Using the labelling given in figure A.3 the triangular basis is:
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
207
50
100
150
200
500 100 150 200
Figure A.2: Mass m atrix for the continuous basis for quadrilaterals ( N = 15).
Figure A.3: Vertex labelling for the standard triangular region
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
208
Vertex A : ( V ) (V )
Vertex B : ( l± s ) (1=4)
Vertex C : (*§^) i
Edge AB : (l= e) (l±e)/> W (0) (l=S)i+l (0 < i < A )
Edge AC : (fa s) (l= i) ( l f i ) p /J p ) (0 < j < P2)
Edge BC : ( l± t ) ( l f t ) ( l i t ) pM (6) ( 0 < i < A ) ,
Interior : ( l= t ) (1 ft) p W (0) ( l f t ) ‘ (1 ft) P * ? ' \ b )
(0 < i, j; i < P i ; i + j < P2, Pi < P2)■
The mass matrix (<&, <f>j) is shown graphically in A.4. In this figure we have
ordered the modes in the geometric sense, i.e. vertex modes then edge modes then
interior modes last. We see that the interior-interior matrix is strongly banded.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
209
Figure A.4: Mass m atrix for the continuous basis for triangles (N = 15).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
A .2.2 T h ree-D im en sion a l E xpansions210
a) H exahedral Domain b) Prismatic Dom ain
c) Pyram idic Domain d) Tetrahedral Dom ain
Figure A.5: Vertex labelling for three-dimensional standard regions
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
H exahedral D om ain
Using the labelling given in figure A.5 (a) the hexahedral basis is:
Vertex A : (1=*) ( t o ) ( t o ) Vertex B : ( t o ) ( t o ) ( t o )
Vertex C : ( t o ) ( t o ) ( t o ) Vertex D : ( t o ) ( t o ) ( t o )
Vertex E : ( t o ) ( iz i ) ( t o ) Vertex F : ( t o ) ( t o ) ( t o )
Vertex G : ( l = £ ) ( t o ) ( t o ) Vertex H : ( t o ) ( t o ) ( t o ) ,
Edge AB : (»=*) ( t o ) ( t o ) ( t o ) (0 < i < P l)
Edge CD : ( t o ) ( t o ) p u (a) ( t o ) ( t o ) :
Edge EF : ( t o ) ( t o ) (o) ( t o ) ( t o )
Edge GH : ( t o ) ( t o ) ^ t ( a ) ( t o ) ( t o ) i
Edge AC : ( ¥ ) ( ¥ ) ( ¥ ) Pi - d b) ( ¥ ) (o < j < p l)
Edge BD ■■ ( ¥ ) ( ¥ ) ( ¥ ) p }-\(b) ( ¥ )
Edge EG : ( t o ) ( t o ) ( t o ) p V (j) ( t o ) :
Edge FH : ( t o ) ( t o ) ( t o ) p U (&) ( t o ) :
Edge AE : ( t o ) ( t o ) ( t o ) ( t o ) pW l(c) (o < k < p 3)
Edge BF : ( t o ) ( t o ) ( t o ) ( t o ) pM l(c) :
Edge CG : ( t o ) ( t o ) ( t o ) ( t o ) p tU (C) :
Edge DH : ( t o ) ( t o ) ( t o ) ( t o ) p « l(c) :
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
212
Face ABCD : (1=*) ( i f* ) P h \(a) ( i f ) ( M ) P b\(b) ( i f )
Face EFGH : ( i f ) (l±») * ( » ) (If* ) (1±S) P / i ( 6 ) ( i f s )
(0 < i, j ; * < P i , j < P2)
Face ABEF : ( i f ) (I f* ) P « ( 0) (1=4) (If*) ( i f ) F " ( c )
Face CDGH : (1=*) { i f ) P " ( a ) ( lf4 ) { i f ) ( if* ) P & ( c)
(0 < i, A;; i < Pi; k < P3)
Face ACEG : ( 1 ^ ) (1 ^ ) (1±4) p U (j) (1 ^ ) (If*) P « l(c)
Face BDFH : (l±a) (1=4) (1±») p / ' \ (4) (I js ) (J±s) pWl(c)
(0 < j , k; j < P2; k < P3)
Interior : { i f ) ( I f ) P & W ( f ) ( i f ) P & ( b ) { f ) ( f ) P " ( c )
(0 < i , j , k; i < P i ; j < P 2; k < P 3).
Prism atic D om ain
Using the labelling given in figure A.5(b) the prismatic bases is
( 1 - c )
Vertex A : (1=*) (1=4) (1=*) Vertex B : ( i f ) ( i f ) ( f )
Vertex C : ( i f ) ( i f ) ( i f ) Vertex D : ( if* ) ( i f ) ( f )
Vertex E : (±=a) (±±s) Vertex F : (±±a) (±±s)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Reproduced withpeimjssion of (he copyright owner. Further reproduce
reproduction prohibited without permission.
214
Edge AB : ( t o ) ( t o ) (a) ( t o ) ( t o ) (0 < i < Pt )
Edge CD : ( t o ) ( t o ) f " ( a) ( t o ) ( t o ) i
EdgeEF : ( t o ) ( t o ) P # (a) ( t o ) !
Edge AC : ( V ) ( ¥ ) ( ¥ ) ( # ' ( .0 < j< P 2)
Edge BD : ( t o ) (1=4) ( t o ) p W (6) (!? )I+1 :
EdgeA E : ( t o ) ( t o ) ( t o ) ( t o ) P U l(c) (0 < fc < P3)
Edge BF : ( t o ) ( t o ) ( t o ) ( t o ) P u , (c) :
EdgeCE : ( t o ) ( t o ) ( t o ) ( t o ) P « ( C) :
Edge DF : ( t o ) ( t o ) ( t o ) ( t o ) :
Face ABCD : ( t o ) ( t o ) P t\(a ) ( t o ) ( t o ) p f y b ) ( t o ) J+1
(0 < i , j ; i < PX’j < P2)
Face ABEF : ( t o ) ( t o ) P u (o) ( t o ) ( t o ) ( t o ) PjU (c)
Face CDEF : ( t o ) ( t o ) P M (o) ( t o ) ( t o ) ( t o ) f u (c)
(0 < i, k\ i < P i;k < P3)
Face ACE : ( t o ) ( t o ) ( t o ) P M,(6) ( t o ) ’+1 ( t o ) P ^ \ c)
Face BDF : ( t o ) ( t o ) ( t o ) ^ l( j) ( t o ) >+1 ( t o ) P V ? - \c )
(0 < j , k] j < P2; j + k < P3; p 2 < p 3)
Interior :
(0 < i , j , k; i < Px- j < P2; j + k < P3; P2 < P3.)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
215
Figure A.7: Mass m atrix for the continuous basis for prisms (N = 15).
P yram idic D om ain
Using the labelling given in figure A. 5(c) the pyramidic basis is :
r = 2 i i ± 4 - i , s = 2 ^ 4 - i ,(1 - c ) ( 1 - c )
t = c.
Vertex A : ( fca ) ( i j i ) (l=s) Vertex B : (1±») (i=S) (l=s)
Vertex C : ( V ) ( ¥ ) ( ¥ ) Vert<=* D : ( ¥ ) (* ? )
Vertex E :
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
216
Edge AB : ( 1 f t ) ( i f a ) p h \ ( a) (1=1) (i= s)i+1 (0 < i < P t)
Edge CD : (i= e) (I± s) ^ \ {a) (i± i) ( I je y +1 :
Edge AC : (1=*) (l= i) (1±4) P " ( b ) ( i= t) i+l (0 < J < P2)
Edge BD : (If®) (1=4) ( l± t) p £ ( b) ( t s ) * 1 I
EdgeA E : ( I f* ) (1=4) (1 ft) (if* ) f « ( c) (0 < k < P3)
Edge BE : (1±S) ( l= i) (J=£) (l±e) p u i(c) :
EdgeCE : ( I f a ) ( l f i ) (1=«) (J±s) pW l(c) :
EdgeD E : (l± a) (1±S) (l=e) (i±£) P Ml(c) :
Face ABCD : (1=*) O f* ) P & ifi) ( i f t ) (1 ft) p U (4) (i= *)i+)+1
(0 < 2, i ; 2 < P u j < P2)
Face ABE : ( i f* ) ( if* ) J * ‘ (a ) (1=*) (l=*)i+1 ( if* ) P “ « . ‘(c)
Face CDE : ( i f* ) ( if* ) P « (a) (1 ft) ( if* ) i+I ( if* ) P ^ ( c)
(0 < i, k; i < P i; i + k < P3; Pi < P3)
Face ACE : ( i f* ) ( i f t ) ( l± t) ph \(b ) ( i f * ) * 1 ( if* ) ^ . ‘(c)
Face BDE : ( I f* ) (1=4) (1 ft) P « (4) (i=*)J+1 ( i p ) P?_*'-\c)
(0 < j , k; j < P2- j + k < P3; P2 < P3)
Interior :
(V) (‘f2) **.(«) (¥) (¥) pm (*?)**" Of*) p™lw(c) (0 < i , j , k; i < Px; j < P2; i + j + k < P3; P L, P2 < P3.)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
217T e tra h e d ra l D o m a in
Using the labelling given in figure A.5(d) the tetrahedral basis is
(l + q) 2 (1 + 6) 1 t =( - b - c ) ’ (1 — c) ’
Vertex A : ( i p ) ( i p ) ( i p ) Vertex B : ( i p ) ( i p ) ( I p )
Vertex C : ( I p ) ( i p ) ( i p ) Vertex D : ( i p )
Edge AB : ( i p ) ( i p ) i ^ ) ( i p ) '+I ( i p ) i+1 (0 < i < f t )
Edge AC : ( I p ) ( i p ) ( i p ) p / i f t ) ( i p ) J+1 (0 < j < P2)
Edge BC : ( I p ) ( i p ) ( I p ) pM l(6) ( i p ) J+1 :
Edge AD : ( I p ) ( I p ) ( I p ) ( I p ) P “ (C) ( 0 < k < P 3)
Edge BD : ( I p ) ( I p ) ( I p ) ( I p ) Pt‘i ( c ) I
Edge CD : ( I p ) ( i p ) ( i p ) pW l(c) ;
Face ABC : ( I p ) ( i p ) P « ( a ) ( i p ) i+I ( I p ) p “ +U(i,) ( l p ) ' ^ +1
(0 < i , j; i < Pi; i + j < P2\ Pi < P2)
Face ABE : ( I p ) ( I p ) f £ i ( a ) ( lp ) * +‘ ( I p ) ‘+1 ( I p )
(0 < i, fc; i < P i \ i + k < P 3; P\ < P3)
Face ACE : (-
Face BCE :
(0 < j, k ; j < P2; j + k < P3; P2 < P3)
(0 < i, k ; i < Pi, i + k < P 3; P i <
( i p ) ( i p ) ( i p ) p ji\(b ) ( i p ) J+1 ( i p ) i f 5 u (c)
( Ip ) ( I p ) ( I p ) Pjl\(.b) ( i p ) ’+1 ( i p ) p £ p ‘(c)
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
218
Interior :
0 ? ) (* ?) ( W * 1 ( ¥ ) p?~t ww (i? ) '+i+111? ) pI-? * 1' \c)(0 < i j , k ; i < Px; i + j < P 2; i + j + k < P3; Px < P2 < P3.)
3K3
I
100
200
300 -
400
500
GOO
300 400 500 600 700
Figure A.8: Mass m atrix for the continuous basis for tetrahedra (N = 15).
A.3 Nodal Basis
A .3.1 T w o-D im ension al E xpansions
In the following section we define the notation hp (£) to be the Lagrange polynomial of
order P through the ( P + l ) zeros of the Guass-Lobatto-Legendre integration scheme.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
219
(b)(a)
Figure A.9: Vertex labelling for the (a) standard quadrilateral and (b) standard Hexahedral region
Q uadrilateral D om ain
Using the labelling given in figure A.9(a) the quadrilateral bases is:
Vertex A : /if1 (a)/if2 (6)
Vertex B : /i%(a)/if2(6)
Vertex C : /if1 (a) hp2 (b)
Vertex D : hpl(a)hp2(b)
Edge AB : /if1 (a)/if2(b) (0 < i < Pi)
Edge CD : /if1 (a)/if2 (6) (0 < i < Px)
Edge AC : /if1 (a)/if2 (6) (0 < j < P2)
Edge BD : /if[(a)/if2(6) (0 < j < P2),
Interior : /if1 (a)/if2(6) (0 < i ,j; i < Px; j < P2).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
220A .3.2 T hree-D im en sional E xpan sion s
H exahedral D om ain
Using the labelling given in figure A.9(b) the hexahedral basis is:
Vertex A : hPi (a) hP2 (b) hPz (c) Vertex B : hp\(a)hP2 (b)hPz {c)
Vertex C : hPl (a)hPp2(b)hPz(c) Vertex D : hp\(a)hPl(b)hP3(c)
Vertex E : hPl {a)hP2 (b)hp3 (c) Vertex F : hp[(a)h^2(6)/ip^(c)
Vertex G : hPl (a) hpl (b) hp* (c) Vertex H : hp\ (a) hp2 (6) hPpz (c)
Edge AB : h f l (a) Hq2 (b) hPz (c) (0 < * < Px)
Edge CD : h f l (a)h£2(b)h£3(c) :
Edge EF : hp'(a )hP2{b)h%{c) ;
Edge GH - h f l (a )h % m p l(c ) :
Edge AC : hPl (a) hPi (6) hPs (c) ( 0 < j < P 2)
Edge BD : hPl (a) hPi {b)hPz (c) :
Edge EG : hp'{a)hP2{b)h%{c) :
Edge FH : tf (a )/» ? (6 )A § (c )
Edge AE : hp'(a )hP2(b)hPz(c) (0 < k < Pz)
Edge BF : hp (a)hP2 {b)hPz (c) '■
Edge CG : h ^(a )h ^ l(b )h ^(c ) \
Edge DH : hp\{a)h%{b)hPz{c) :
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Face ABCD : /if1 (<z)/if2 (6)/if3 (c)
Face EFGH : h[l (a)hf2(b)h^{c)
(0 < i , j ; i < Px; j < P2)
Face ABEF : /if1 (a) /if2 (b) /if3 (c)
Face CDGH : /if1 (a)/if2 (6)/if3 (c)
(0 < i, k; i < P i ,k < P3)
Face ACEG : h ? (a )h f2(b)h?(c)
Face CDGH : /ig (a )/if2(6)/if3(c)
(0 < j, k; j < P i ,k < P3)
Interior : /if1 (a) /if2 (6) /if3 (c)
(0 < i, j , k; i < Pi; j < P2;k < P3).
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Bibliography
[1] Karypis G. and Kum ar V. Multilevel k-way Partitioning Scheme for Irregu
lar Graphs. Technical report, Department of Computer Science, University of
Minnesota, 1995.
[2] Tufo H.M. Generalized communication algorithms for parallel fern codes. Tech
nical report, Center for Fluid Mechanics, Brown University, 1997.
[3] Peiro J., Peraire J., and Morgan K. 2D Felisa System - Reference Manual. Dept,
of Aeronautics, Imperial College, 1994.
[4] Barth T.J. Steiner trianguiization for isotropic and stretched elements. AIAA-
95-0213, 1995.
[5] Babuska I., Szabo B.A., and Katz I.N. The p-version of the finite element
method. SIA M J. Numer. Anal., 18:515, 1981.
[6] Babuska I. and Dorr M. R. Error estimates for the combined h- and p-version
of the finite element m ethod. Num. Math., 37:257, 1981.
222
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
223[7] Patera A. T. A spectral method for fluid dynamics: Laminar flow in a channel
expansion. J. Comp. Phys., 54:468, 1984.
[8] Kamiadakis G.E., Bullister E.T., and Patera A.T. A spectral element method for
solution of two- and three-dimensional time dependent Navier-Stokes equations.
In Finite Element Methods for Nonlinear Problems, Springer-Verlag, page 803,
1985.
[9] Oden J. T. Optimal hp-finite element methods. Technical Report TICAM Report
92-09, University of Texas a t Austin, 1992.
[10] Bemardi C., Maday Y., and Patera A. T. A new conforming approach to domain
decomposition: the m ortar element method. In Nonlinear Partial Differential
Equations and Their Applications, Pitman and Wiley, 1992.
[11] Biswas R., Devine K., and Flaherty J. Parallel, adaptive finite element methods
for conservation laws. Appl. Num. Math., 14:255-283, 1994.
[12] Henderson R. D. Details of the drag curve near the onset of vortex shedding.
Phys. Fluids, 7(9):l-3, 1995.
[13] Szabo B. and Babuska I. Finite Element Analysis. John Wiley & Sons, 1991.
[14] Zienkiewicz 0 . C. and Taylor R. L. The Finite Element Method: vol. 1. McGraw-
Hill, 1989.
[15] Peano A. Hierarchies of conforming finite elements for plane elasticity and plate
bending. Comp. & Maths, with Appls., 2:221, 1976.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
224[16] Dubiner M. Spectral methods on triangles and other domains. J. Sci. Comp.,
6:345, 1991.
[17] Proriol J. Sur une Famille de Polynomes a deux Variables Orthogonaux dans un
Triangle. C.R. Acad. Sci. Paris 257:2459-2461, 1957.
[18] Koomwinder T. Two-variable Analogues of the Classical Orthogonal Polyno
mials Theory and Applications o f Special Functions, Askey R.A. ed., Academic
Press, 1975.
[19] G. W. Zumbusch Symmetric Hierarchical Polynomials and the Adaptive h-p-
Version. Proc. of the Third Int. Conf. on Spectral and High Order Methods,
ICOSAHOM’95, Houston, TX, 1995.
[20] Sherwin S. J. and Kamiadakis G. E. A triangular spectral element method;
applications to the incompressible Navier-Stokes equations. Comp. Meth. Appl.
Mech. Eng., 123:189-229. 1995.
[21] Sherwin S. J. and Kamiadakis G. E. A new triangular and tetrahedral basis
for high-order finite element methods. Int. J. Num. Meth. Eng., 123:3775-3802
1995.
[22] Aftosmis M., Melton J., and Berger M. Adaptation and surface modeling of
cartesian mesh methods. In AIAA-95-1725, 1995.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
225[23] Gordon W .J., and Hall C.A. Transfinite Element Methods: Blending Function
Interpolation Over Arbitrary Curved Element Domains. Num. Math., 21:109
1973.
[24] Parthasarathy V., Kallinderis Y., and Nakajima K. Hybrid adaption method
and directional viscous multigrid with presm atic-tetrahedral meshes. In AIAA-
95-0670, Reno, NV, 1995.
[25] Mavriplis D. J. and Venkatakrishnan V. A unified multigrid solver for the Navier-
Stokes equations on mixed element meshes. In AIAA-95-1666, San Diego, CA,
1995.
[26] Tafli D., Soetrisno M., and Imlay S. A parallel algorithm for hybrid structured-
unstructured grids. In AIAA-96-0287, Reno, NV, 1996.
[27] Sherwin S. J. Hierarchical hp finite elements in hybrid domains. Finite Elements
in Analysis and Design, 27:109—119, 1997.
[28] Wingate B.A. and Taylor M.A. The natural function space for triangular and
tetrahedral spectral elements. 1998. subm itted to SIAM J. Num. Anal.
[29] Owens R.G. Spectral approximations on the triangle. Subm itted Proc. R. Sec.
Lond. A, 1997.
[30] Sherwin S. J. Triangular and Tetrahedral spectral/hp finite element methods for
fluid dynamics. PhD thesis, Princeton University, 1995.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
226[31] Kamiadakis G. E. and Sherwin S. J. Spectral hp Element Methods fo r CFD.
Oxford University Press, In Press.
[32] W ingate B. A. and Boyd J. P. Triangular spectral element methods for geophys
ical fluid dynamics applications. Houston J. o f Mathematics, page 305, 1996.
[33] R.0nquist E.M. Optimal spectral element methods fo r the unsteady three-
dimensional incompressible Navier-Stokes equations. PhD thesis, Massachusetts
Institute of Technology, 1988.
[34] Mavriplis C. and Rosendale J.V. Triangular spectral elements for imcompressible
fluid flow. Number AIAA-93-3346-CP, pages 540-545, 1993.
[35] Chen Q. and Babuska I. Approximate optimal points for polynomial interpola
tion of real functions in an interval and in a triangle. Comp. Meth. Appl. Mech.
Eng., 128:405, 1995.
[36] Hesthaven J.S. and Gottlieb D. Stable spectral methods for conservation laws
on triangles w ith unstructured grids, submitted SIAM J. Numer. Anal., 1997.
[37] Hesthaven J.S. Private Communications, 1998.
[38] Cuthill E.H. and McKee J. Reducing the bandwidth of sparse symmetric systems.
In Proc. 24th Nat. Conf. Assoc. Comp. Mach., page 157. ACM Publications,
1969.
[39] Kamiadakis G. E., Israeli M., and Orszag S.A. High-order splitting methods for
the incompressible Navier-Stokes equations. J. Comp. Phys., 97:414, 1991.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
227[40] Henderson R. D. Unstructured Spectral Element Methods: Parallel Algorithms
and Simulations. PhD thesis, Princeton University, 1994.
[41] Wannier G.H. A contribution to the hydrodynamics of lubrication. Q. Appl.
Math., 8:1, 1950.
[42] Kovasznay L.I.G. Laminar flow behind a two-dimensional grid. Proc. Cambridge
Philos. Soc., page 44, 1948.
[43] Williamson C.H.K. Oblique and parallel modes of vortex shedding in the wake
of a circular cylinder at low reynolds numbers. J. Fluid Mech., 206:579, 1989.
[44] Hammanche M. and Gharib M. An experimental study of the parallel and oblique
shedding in the wake from circular cylinders. J. Fluid Mech., 232:567, 1991
[45] Ma Xia and Kamiadakis G.E. Dynamics and low-dimensionality of the turbulent
near-wake. J. Fluid Mech., 1998. submitted.
[46] Evangelinos C., Sherwin S. J., and Kamiadakis G.E. Parallel DNS algorithms
on unstructured grids. ?, 1998. submitted.
[47] Evangelinos C. Parallel Spectral/hp Methods and Simulations of Flow/Structure
Interactions. PhD thesis, Brown University, 1999.
[48] Newman D. A Computational Study of Fluid/Structure Interactions: Flow-
Induced Vibrations o f a Flexible Cable. PhD thesis, Princeton University, 1996.
[49] Hirt C.W., Amsden A.A., and Cook H.K. An arbitrary Lagrangian-Eulerian
computing method for all flow speeds. J. Comp. Phys., 14:27, 1974.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
228[50] Hughes T.J.R., Liu W .K., and Zimmerman T.K. Lagrangian-Eulerian finite
element formulation for incompressible viscous flows. Comput. Methods Appl.
Mech. Eng., 29:329, 1981.
[51] Donea J., Giuliani S., and Halleux J.P. An arbitrary Lagrangian-Eulerian fi
nite element method for transient dynamic fluid-structure interactions. Comput.
Methods Appl. Mech. Eng., 33:689, 1982.
[52] Nomura T. and Hughes T. J.H. An arbitrary Lagrangian-Eulerian finite element
method for interaction of fluid and a rigid body. Comput. Methods Appl. Mech.
Eng., 95:115, 1992.
[53] Lohner R. and Yang C. Improved ale mesh velocities for moving bodies. Comm.
Num. Meth. Eng. Phys., 12:599-608, 1996.
[54] Ho L.-W. A Legendre spectral element method fo r simulation of incompressible
unsteady free-surface flows. PhD thesis, Massachustts Institute of Technology,
1989.
[55] Kamiadakis G.E., Israeli M., and Orszag S.A. High-order splitting methods for
the incompressible Navier-Stokes equations. J. Comp. Phys., 97:414, 1991.
[56] Beskok A. and W arburton T.C. Micro-fluidic design and fluid-structure interac
tion analysis of a micro-pump. IMECE 98, Proceedings o f the ASM E Dynamic
Systems and Control Division, 1998. submitted.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
229[57] Van Lintel H.T.G., Van de Pol F.C.M., and Bouwstra S. A piezoelectric microp-
ump based on micromachining of silicon. Sensors and Actuators, 15:153-167,
1988.
[58] Esashi M., Shoji S., and Nakano A. Normally closed microvalve fabricated on a
silicon wafer. Sensors and Actuators, 20:163—169, 1989.
[59] Smits J.G . Piezoelectric micropump with three valves working peristaltically.
Sensors and Actuators A, 21-23:203-206, 1990.
[60] Message Passing Interface Forum MPI: A Message-Passing Interface Standard
Technical report, June 1995.
[61] Zachary A.L., Malagoli A., and Colella P. A higher-order godunov method for
multidimensional ideal magnetohydrodynamics. SIA M J. Sci. Stat. Comp., 15,
1994.
[62] Powell K.G. An approximate Riemann solver for magnetohydrodynamics (that
works in more than one dimension). Technical Report ICASE Report 94-24,
ICASE, NASA Langley, 1994.
[63] Powell K.G., Roe P.L., Myong R.S., Gombosi T., and De Zeeuw D. An Upwind
Scheme for Magnetohydrodynamics. Proc. of A IA A 12th Computational Fluid
Dynamics Conference, 661, AIAA , 1995.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
230[64] Dahlburg R.B. and Picone J.M. Evolution of the orszag-tang vortex system in
a compressible medium, i. initial average subsonic flow. Phys. Fluids B, 1(11),
1989.
[65] Brackbill J.U. and Barnes D.C. The effect of nonzero V - b on the numerical
solution of the magnetohydrodynamic equations. J. Comp. Phys., 35:426-430,
1980.
[66] Orszag S.A. and Tang C. Small-scale structure of two-dimensional magnetohy
drodynamic turbulence. J. Fluid Mech., 90(1):129—143, 1979.
[67] Cockbum B. and Shu C.-W. The local discontinuous Galerkin for time dependent
convection-diffusion systems. SIAM J. Num. Anal., 1997. submitted.
[68] Cockbum B. and Shu C.-W. TVB Runge-Kutta local projection discontinuous
Galerkin finite element method for conservation laws II : General framework .
Math. Comp., 52:411-435, 1989.
[69] Cockbum B., Lin S.Y.,and Shu C.-W. TVB Runge-Kutta local projection discon
tinuous Galerkin finite element method for conservation laws III: one dimensional
systems. J. of Comp. Phys., 84:90-113, 1989.
[70] Cockbum B., Hou S., and Shu C.-W. The Runge-Kutta local projection dis
continuous Galerkin finite element method for conservation laws IV: The multi
dimensional case. J. Comp. Phys., 54:545, 1990.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
231[71] Cockbum B. and Shu C.-W. The Runge-Kutta discontinuous Galerkin m ethod
for conservation laws V: multidimensional systems. J. of Comp. Phys., 141:199-
224, 1998.
[72] Lomtev I. and Kamiadakis G. E. Simulations of viscous supersonic flows on
unstructured meshes. AIAA-97-0754, 1997.
[73] Lomtev I., Quillen C.B., and Kamiadakis G. E. Spectral/hp methods for viscous
compressible flows on unstructured 2d meshes. J. Comp. Phys., in press, 1998.
[74] Lomtev I. and Kamiadakis G. E. A discontinuous Galerkin method for the
Navier-Stokes equations. Int. J. Num. Meth. Fluids, submitted, 1997.
[75] Baumann C.E. and Oden J. T. A discontinuous hp finite element method for
the solution of the Euler equations of Gas Dynamics. In Tenth International
Conference on Finite Elements in Fluids, Jan. 5-8 1998.
[76] Baumann C.E. and Oden J. T. A discontinuous hp finite element method for
the solution of the Navier-Stokes equations. In Tenth International Conference
on Finite Elements in Fluids, Jan. 5-8 1998.
[77] Baumann C.E. and Oden J. T. The Discontinuous Galerkin Method Applied
to CFD Problems. In SIA M 45th Anniversary Meeting, High order methods for
compressible flow calculations, July 14-18 1997.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
232[78] Oden J. T. , Babuska I., and Baumann C.E. A discontinuous hp finite element
method for diffusion problems. TICAM Report 97-21. J. of Comp. Phys. - Re
vision., 1997.
[79] Babuska I., Oden J. T. , and Baumann C.E. A discontinuous hp finite element
method for diffusion problems: 1-D Analysis . TICAM Report 97-22. Computer
and Mathematics w ith Applications - Revision., 1997.
[80] Lomtev I. Spectral/hp Methods for Viscous Compressible Flows on Unstructured
Meshes. PhD thesis, Brown University, 1999.
[81] Beskok A. Simulations and Models fo r Gas Flows in Microgeometries. Ph.D.
Thesis, Princeton University, Princeton, NJ, June 1996.
[82] Kirby R. M. PhD thesis: work in progress. Brown University, 1999.
[83] Barth T. J. SIM PLEX2D User Guide. Advanced Algorithms and Applications
Branch, NASA Ames Research Center, 1995.
[84] Priest E. Solar Magnetohydrodynamics. D. Reidel Publishing Company, 1982.
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
%
V*
IMAGE EVALUATION TEST TARGET (Q A -3 )
✓
A
/ -
1.0
l.l
1.25
2.8
3.2
in
1.4
IM12.2
2.0
1.8
1.6
150mm
A P P L I E D A IM/4GE . Inc1653 East Main Street
- = = ~- Rochester, NY 14609 USA Phone: 716/482-0300
- = ’-: = ‘ Fax: 716/288-5989
0 1993. Applied Image. Inc.. All Rights R eserved
permission of the copyright owner. Further reproduction prohibited without permission.