developments of a discrete adjoint structural solver for ...marc schwalbach1,tom verstraete2,...
TRANSCRIPT
-
Developments of a Discrete Adjoint Structural Solver for Shapeand Composite Material Optimization
Marc Schwalbach1 ,Tom Verstraete2, Nicolas R. Gauger3
AD 2016September 15, 2016
Oxford, UK
1von Karman Institute for Fluid Dynamics2Queen Mary University of London3Technische Universität Kaiserslautern
-
I C++
I Eigen, PETSc, SLEPc
I AD tool CoDiPack
2 a
-
I FEM linear stress analysis
I FEM vibration analysis
I extension to composite (shell) elements
3 a
-
linear stress analysis
1. input x ∈ Rn
2. linear system solve A(x)u = b(x)
3. output σmax(u(x)) ∈ R
x(1) =∂σmax∂x
?
4 a
-
5 a
-
6 a
-
b(1) = A−Tu(1)
A(1),i,j = −ujb(1),i
7 a
-
Ab(1) = u(1)
A(1),i,j = −ujb(1),i
8 a
-
9 a
-
rotating beam. ‖ ∂σmax∂x
‖ ∈ R90
10 a
-
11 a
-
DOFs ≈ 30k ≈ 200k 200k, RealReverseIndexrelative run time 2.5× T 2.1× T 2.8× Trelative peak memory 7.7×M 6.4×M 5.8×M
gradient evaluation (forward + reverse run). M : peak memory of primal. T : run time of primal
12 a
-
run time breakdown. 200k DOF axial fan. primal reverse AD
13 a
-
14 a
-
15 a
-
vibration analysis
1. assemble mass matrix B(x)
2. solve generalized eigenvalue system: (A− λiB)ui = 0
x(1) =∂λi∂x
?
16 a
-
17 a
-
A(1),ij = λ(1),kuk,iuk,j (1)
B(1),ij = −λ(1),kλkuk,iuk,j (2)
18 a
-
19 a
-
rotating beam. ‖ ∂λ0∂x‖ ∈ R90
20 a
-
21 a
-
run time breakdown. 200k DOF axial fan. primal reverse AD
22 a
-
I same linear and eigenvalue solver
I extended design variables x∗ = (x,V )
I lamination params
V = V (t,θ)
24 a
-
I different stiffness and mass matrices Ac(x∗), Bc(x
∗)
I different failure criterion σTW
∂σTW∂x∗
,∂λk∂x∗
25 a
-
26 a
-
27 a
-
composite flat plate. eigenmode 0 eigenmode 3
28 a
-
∂λ0∂x
∂λ3∂x
29 a
-
∂λ0∂V11
30 a
-
∂λ0∂V11
∂λ3∂V11
31 a
-
I AD’d linear stress analysis, separate linear solver
I AD’d vibration analysis, separate EV solver
I extension to AD’d composites
32 a
-
Developments of a Discrete AdjointStructural Solver for Shape andComposite Material Optimization
Marc Schwalbach1
Tom Verstraete2
Nicolas R. Gauger3
financial support
European Commission - IODA4project
1von Karman Institute for Fluid Dynamics2Queen Mary University of London3Technische Universität Kaiserslautern4Industrial Optimal Design using Adjoint CFD 32 a