improving spray models for · pdf fileimproving spray models for advancedadvanced combustion...

IIMPROVINGMPROVING SPRAYSPRAY MODELSMODELS FORFORIIMPROVINGMPROVING SPRAYSPRAY MODELSMODELS FORFORADVANCEDADVANCED COMBUSTIONCOMBUSTION STRATEGIESSTRATEGIES

Federico Perinia, Paul C. Milesb, Rolf D. Reitza

aUniversity of Wisconsin–MadisonUniversity of Wisconsin MadisonbSandia National Laboratories

IEA 36th Task Leaders Meeting 2014

Acknowledgements

g“Energy Conservation and Emission Reduction in Combustion”

Stavanger (Norway), June 9-13, 2014

AcknowledgementsD.O.E. Office of Vehicle Technologies, P.M. Gupreet Singh

Sandia National Laboratories

Engine Research Center, University of Wisconsin – MadisonIEA Combustion 2014 – Stavanger, Norway – June 9-13, 2014

1

TransitionTransition toto largelarge--scalescale computationscomputationsU d t d th ‘l lit ff t ’ f fl iti l d th l- Understand the ‘locality effects’ of flow, compositional and thermal

non-uniformities on combustion- Prepare the path towards comprehensive flow and transport modelling(LES DNS) and future engine studies with full engine geometry

First KIVA4 implementation (Torres, 2006) as the base code

(LES, DNS) and future engine studies with full engine geometry

p ( , ) Need for a framework that is tailored to internal combustion engine simulations Buggy but enabled support for unstructured geometries

Large-scale combustion chemistry Large scale combustion chemistry Sparse Analytical Jacobian chemistry solver (‘SpeedCHEM’) High-Dimensional cell Clustering

Extended and improved spray modelling

Parallelization of the flow field and spray solution


2

PPC PPC mixturemixture preparationpreparation experimentsexperimentsExperiments carried out at Sandia National pLaboratories by P.C. Miles, D. Sahoo, S. Busch

- Optically-accessible Sandia-GM 1.9L engineBosch CRI2 2 7 hole injector- Bosch CRI2.2 7-hole injector

- Variable swirl ratio intake: Rs = 1.5 to 4.5- Fuel for mixture preparation studies: PRF25

PLIF equivalence ratio measurements- PLIF equivalence ratio measurements

CA = -15 deg ATDC0 5

CA = -15 deg ATDC0 5

SAE 2013-24-0061

30

35PPC injection rates, 8.8 mg, CRIP 2.2

pinj = 500 bar

b

e fr

om fi

rede

ck [c

m]

-1.0

-0.5

0.0

0.5

P1P2

P3

e fr

om fi

rede

ck [c

m]

-1.0

-0.5

0.0

0.5

P1P2

P3 15

20

25

30

rate

[g/s

]

pinj = 860 bar

pinj = 1220 bar

r axis [cm]

dist

ance

0 1 2 3 4

-2.0

-1.5

0 0.5 1 1.5 2 2.5 3 3.5 4

r axis [cm]

dist

ance

0 1 2 3 4

-2.0

-1.5

0 0.5 1 1.5 2 2.5 3 3.5 40 0 2 0 4 0 6 0 8 1

0

5

10inj.


3

0 0.2 0.4 0.6 0.8 1x 10-3time [s]

PPC PPC mixturemixture preparationpreparation validationvalidation30

/s]

pinj=860bar RS=2.2 SOI=-23.3

t hP1 (squish)CA 10 0 0

10

20

inje

ctio

n ra

te [c

msector meshSquish mixing = slightly

underestimated

1.5

CA = -10.0experiment simulation

-20 -15 -10 -50

crank angle [deg aTDC]

11

0.5


4 0phi [-]


/s]


t hP2 (rim)

CA 10 0 0

10

20

inje

ctio

n ra

te [c

msector meshA few millimeters close tothe bowl rim not availablein the experimental image

1.5


-20 -15 -10 -50


11

0.5


5 0phi [-]


/s]


t hP3 (bowl)CA 10 0 0

10

20

inje

ctio

n ra

te [c

msector meshSlight underprediction of jet penetration reachingback up to the centerline

1.5


-20 -15 -10 -50


11

0.5

Reasonably accurate equivalence ratio


6 0phi [-]

equivalence ratio prediction where

measured data available

Spray Spray modellingmodelling improvementimprovement efforteffort

Current spray model validation(unstructured code, sector mesh same to SAE2013-01-1105) Current spray model validation(unstructured code, sector mesh same to SAE2013-01-1105)

Flow prediction validation(unstructured full engine model generalized RNG k-ɛ closure)

Flow prediction validation(unstructured full engine model generalized RNG k-ɛ closure)(unstructured full engine model, generalized RNG k-ɛ closure)

(CaF 2013, submitted; THIESEL 2014, submitted) (unstructured full engine model, generalized RNG k-ɛ closure)

(CaF 2013, submitted; THIESEL 2014, submitted)

Spray model improvement and calibration(Sandia ECN spray experiments) Spray model improvement and calibration(Sandia ECN spray experiments)

Full engine cycle validation, Sandia 1.9L light duty engine(unstructured, full mesh and full cycle engine calculation)

Spray A

Injection-induced turbulenceeffects on near-nozzle flow

Model study to capture near-SOI transient


7

effects on near nozzle flow SOI transient

SandiaSandia Spray A Spray A modelingmodeling900K, 60 bar, tinj = 1.5 ms

• [contour] fuel vapor mass fraction, in the range [0, 0.05]• [green dots] liquid phase distribution projectionj [g ] q p p j

50 us 100 us 300 us

IFP

Sandia

ERC

500 us 800 us 1500 us


8

Spray Spray constantsconstants GA GA studystudyM lti l i t ti d l h d t i l t th ff t f h d l• Multiple interacting spray models hard to isolate the effects of each model

• Hard to validate each of these isolated phenomena against experiments• May be aided by future DNS simulations

A GA optimization to answer these questions:

What parameter regions should we move in? When we used to calibrate the spray constants, how muchwere we tweaking the gas-phase prediction too?were we tweaking the gas-phase prediction too? Is there an optimal calibration set, and, does thisinclude “historically used” values or does it suggest new ones, which better fit the newest and highly confident Sandiaexperiments?


9

Spray Spray constantsconstants GA GA studystudy6 model variables6 model variables

Variable name std value range toRT time constant CRT 1.0 0.05 50.0RT wavelength cnst C 0 1 0 01 10 0RT wavelength cnst. CΛRT 0.1 0.01 10.0KH decay timescale cnst. B1 40.0 10.0 100.0Gas-jet Stokes number St 3.0 0.1 5.0Gas-jet entrainment cnst K 0 7 (ideal=0 45) 0 3 3 0Gas-jet entrainment cnst. Kentr 0.7 (ideal=0.45) 0.3 3.0Max gas-jet velocity frac. γ 0.6 0.2 0.9

5 Spray A objectives5 Spray A objectives

Phenomenon merit1) Vapor penetration integral mean squared error (MSE)2) Vapor dispersion mean integral MSE (future addition)2) Vapor dispersion mean integral MSE (future addition)3) Liquid ramp integral MSE4) Steady liquid region mean peneration error5) penetration stdev error


10

5) penetration stdev error

Spray A Spray A objectivesobjectivesSpray A, C12H26, 900K, O2=0.0, tinj=1.5ms

5

6

7m

]j

Experimentsimulation

1) MSE on predicted vapor penetration

t sim d

vvf

2

exp1

2

3

4

pene

trat

ion

[cm

dvt

f0

exp

p1

Spray A, C12H26, 900K, O2=0.0, tinj=1.5ms

0 0.5 1 1.5 210

-3

0

1

time [s]

1

1.2

1.4

[cm

]

p y , 12 26, , 2 , inj

l

x 10time [s]

0 4

0.6

0.8

uid

pene

trat

ion

4,5: error on mean and stdevof steady liquid length

l

0 0.5 1 1.5 23

0

0.2

0.4

ti [ ]liq

u

;;

exp

exp5

exp

exp4 l

llf

lll

f simsim

l


11

x 10-3time [s]

Spray A Spray A objectivesobjectives

3) liquid penetration on the ramp up

0.8

1

tion

[cm

]

rampt sim d

lll

tf

0

2

exp3

1

0.4

0.6

liqui

d pe

netr

at

ramp lt 0

exp

2) Mean integral vapor dispersion MSE (under development)

1

0 1 2x 10

-4

0

0.2

time [s]

axiszzzz

MSEMSEMSEMSEMSEf 504030202 5

1

0.16

i tr

20 30 40 50mm axial

0.08

0.12

e fr

actio

n [-]

experimentsimulation

z = 30 mm

z

0

0.04

mix

ture


12

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

distance [cm]

ResultsResults: breakup : breakup modelmodel constantsconstants

KH decay time scale constant (B1)- affects the liquid ramp phase ( RT breakup not occurring yet), not the steady-stateq p p ( p g y ), y- Vapor penetration: B1 > 50 we do not want breakup to compensate for turbulence!-Liquid ramp: B1 ϵ [35 – 44] converging to the widely validated B1 = 40

RT model constants (wavelength CλRT, timescale CτRT)

Coordinates the variablesColors merit: the bluer, the betterRT model constants (wavelength CλRT, timescale CτRT)

- Crucial to liquid length prediction, which is “a tiny bit” after RT breakup happens- the RT timescale seems to have optimal values ~ CτRT = 3.7; or 2<CτRT<4

,


13

ResultsResults: breakup : breakup modelmodel constantsconstants

KH decay time scale constant (B1)- affects the liquid ramp phase ( RT breakup not occurring yet), not the steady-stateq p p ( p g y ), y- Vapor penetration: B1 > 50 we do not want breakup to compensate for turbulence!-Liquid ramp: B1 ϵ [35 – 44] converging to the widely validated B1 = 40

RT model constants (wavelength CλRT, timescale CτRT)RT model constants (wavelength CλRT, timescale CτRT)- Crucial to liquid length prediction, which is “a tiny bit” after RT breakup happens- the RT timescale seems to have optimal values ~ CτRT = 3.7; or 2<CτRT<4


14

ResultsResults: gas: gas--jet jet modelmodel constantsconstants

Very definite behavior for vapor penetration and liquid rampy p p q pγ ϵ [0.7 – 0.9] Better to apply most of the effective gas-jet velocityKentr ϵ [0.6 – 0.9] for vapor penetration,

[0.8 – 1.5] for liquid ramp slightly higher than currently used

Stokes ϵ [0.8 – 1.0] Significantly smaller than the value St = 3.0 suggested in (Abani and Reitz, 2008) for steady gas-jet modelling Study of

Stokes number effects


15

Stokes number effects

GA GA optimizationoptimization summarysummary

Confirms complex interactions among the models Suggests that the Stokes number calibration used for steadygas jet modelling is overestimated needs a deeper studygas-jet modelling is overestimated needs a deeper study Confirms standard “historically used” and well validated values(e.g., B1 = 40)

Currently setting up a more comprehensive optimization study:- Large number of individuals and generations- GRNG turbulence validated for vapor penetration- Inclusion of jet dispersion merit as an objectivej p j

Final validated calibration set will be used for grid resolutionstudy


16

s udy

GasGas--jetjet modelmodel StokesStokes numbernumber studystudy

Eddy generation frequency in the modeled gas-jet equal tog j qparticle responsiveness in the

spray jet drops

IV StokesStrouhalxf Assumed to be a model

pflow

StokesStrouhalU

• A noticeably smaller range [0 8 1 0] suggested by the GA optimization

calibration parameter St = 3

• A noticeably smaller range [0.8 – 1.0] suggested by the GA optimization

• The constant value can be replaced by a local estimation, exploiting injection-induced turbulence effects at the nozzle


17

GasGas--jetjet modelmodel StokesStokes numbernumber studystudy• Stokes calculation from RNG k-ɛ predicted integral length scale at the nozzlep g g

104near-nozzle Stokes number, Spray A Initial slope jet-induced turbulence

not yet developed in the nozzle cell

102

St [-

]

stable St = (0.75 0.03)

Reasonable – the CFD model isunderresolved at the nozzle!

Reasonable – the CFD model isunderresolved at the nozzle!

0 0 2 0 4 0 6 0 8 1 1 210-2

100stable St (0.75 0.03)

A very stable region during steady-state injection

0 0.2 0.4 0.6 0.8 1 1.2x 10-3time [s]

1.5

cm]

Spray A, C12H26, 900K, O2=0.0, tinj=1.5ms

high St significantly delayed liquidlength development near SOI

0.5

1

uid

pene

trat

ion

[clength development near SOI

Stokes number crucial to modelling jet-


180 0.5 1 1.5

x 10-3

0

time [s]

liq

ExperimentSimulation, integral-length-based St

induced turbulence effects on droppenetration

GasGas--jetjet modelmodel StokesStokes numbernumber studystudy• Constant Stokes, St = 3.0 (gas-jet model), St = 0.75 (as in the steady part) , (g j ), ( y p )• Variable Stokes, St = (computed at nozzle local cell)

1.4Spray A, C12H26, 900K, O2=0.0, tinj=1.5ms

7Spray A, C12H26, 900K, O2=0.0, tinj=1.5ms

0.8

1

1.2

ratio

n [c

m]

4

5

6

ratio

n [c

m]

0.2

0.4

0.6

liqui

d pe

netr

ExperimentSt = 3.0St = 0.75St = (computed) 1

2

3

vapo

r pe

netr

ExperimentSt = 3.0St = 0.75St = (computed)

0 1 2 3 4x 10

-4

0

time [s]

St = (computed)

0 0.5 1 1.5 2x 10

-3

0

time [s]

St = (computed)

Limited effects over steady-state liquid and vapor penetration

Significantly affects the initial liquidcore development transient

E ll t t h f th i iti l d l t “b ” ith St 0 75


19

Excellent match of the initial development “bump” with St = 0.75

FurtherFurther researchresearch directionsdirectionsSpraySpray Extend GA optimization and carry the improved calibration over to engine simulations

Fluid solution Parallelization for large scale computations Parallelization for large-scale computations The KIVA lesson: a simple (Jacobi!) preconditioner can be very robust, and work very well

(30+ years) if tailored to the problem (= coarse but topology-changing mesh) The most used ILUTP+BiCGStab solver “just works well” possible to improve

d h h l l h f h l l l dpreconditioning the physical relationship of the pressure-velocity coupling is exploited

Chemistry Chemistry solver now scales linearly with problem size (sparse analytical Jacobian ) Need to move to scaling much less-than-linearly with the problem size

adjoint-sensitivity-aided partitioned (ASAP) clustering

Turbulence & Transport We are correctly modeling neither turbulent nor molecular transport Simple, linear isotropic turbulence models fail even over a simple gas jet case but widely

used for engines! Accurate transport modelling diffusion viscosity etc


20

Accurate transport modelling diffusion, viscosity, etc

Acknowledgementsg

D.O.E. Office of Vehicle Technologies, P.M. Gupreet SinghSandia National Laboratories


21

improving spray models for · pdf fileimproving spray models for advancedadvanced combustion...

Documents