Numerical Modeling of HCCI Combustion

Salvador M Aceves Daniel L Flowers J Ray Smith Joel Martinez-Frias Francisco Espinosa-Loza Tim Ross

Bruce Buchholz Nick Killingsworth Tom PiggottJonas Edman Charles Westbrook and William Pitz

Lawrence Livermore National Laboratory

Dennis Assanis Aris Babajimopoulos JY Chen Magnus Christensen John Dec Robert Dibble Russ Durrett Randy Hessel Bengt Johansson

Magnus Sjoberg Dick Steeper Axel Zur Loye

External Contributors

August 24 2006 Work performed under the auspices of the US Department of Energy by University of California Lawrence Livermore National Laboratory under Contract W-7405-ENG-48

Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics

Courtesy of Professor Yuuji Ikeda Kyoto University

The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model

High resolution CFD Lower resolution chemical kinetics simulation (105-106 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

102 mm

SAE Paper 2004-01-1910

(402 in)

Pre

ssur

e k

Pa

Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load

9000

Solid lines experimental

8000 Φ=026 Dashed lines numerical

Φ=024 Φ=022

7000 Φ=020Φ=018Φ=016

6000 Φ=014Φ=012Φ=0105000 Φ=008Φ=006

4000 Φ=004

3000

2000-30 -20 -10 0 10 20 30

crank angle degrees

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics

Courtesy of Professor Yuuji Ikeda Kyoto University

The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model

High resolution CFD Lower resolution chemical kinetics simulation (105-106 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

102 mm

SAE Paper 2004-01-1910

(402 in)

Pre

ssur

e k

Pa

Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load

9000

Solid lines experimental

8000 Φ=026 Dashed lines numerical

Φ=024 Φ=022

7000 Φ=020Φ=018Φ=016

6000 Φ=014Φ=012Φ=0105000 Φ=008Φ=006

4000 Φ=004

3000

2000-30 -20 -10 0 10 20 30

crank angle degrees

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model

High resolution CFD Lower resolution chemical kinetics simulation (105-106 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

102 mm

SAE Paper 2004-01-1910

(402 in)

Pre

ssur

e k

Pa

Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load

9000

Solid lines experimental

8000 Φ=026 Dashed lines numerical

Φ=024 Φ=022

7000 Φ=020Φ=018Φ=016

6000 Φ=014Φ=012Φ=0105000 Φ=008Φ=006

4000 Φ=004

3000

2000-30 -20 -10 0 10 20 30

crank angle degrees

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

102 mm

SAE Paper 2004-01-1910

(402 in)

Pre

ssur

e k

Pa

Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load

9000

Solid lines experimental

8000 Φ=026 Dashed lines numerical

Φ=024 Φ=022

7000 Φ=020Φ=018Φ=016

6000 Φ=014Φ=012Φ=0105000 Φ=008Φ=006

4000 Φ=004

3000

2000-30 -20 -10 0 10 20 30

crank angle degrees

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Pre

ssur

e k

Pa

Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load

9000

Solid lines experimental

8000 Φ=026 Dashed lines numerical

Φ=024 Φ=022

7000 Φ=020Φ=018Φ=016

6000 Φ=014Φ=012Φ=0105000 Φ=008Φ=006

4000 Φ=004

3000

2000-30 -20 -10 0 10 20 30

crank angle degrees

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Multi-zone model makes good predictions of emissions over the full range of operation

004 006 008 010 012 014 016 018 020 022 024 026 028 Equivalence ratio

0

10

20

30

40

50

60

70

80

90

100 Fu

el C

arbo

n in

to e

mis

sion

s (

) Solid lines experimental Dashed lines numerical

CO2 CO HC OHC

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Model can tell us the location in the cylinder where different pollutants originate

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control

PCCI through high EGR that does not PCCI through early mix well with fresh charge (VVT CAI) direct injection

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion

High resolution CFD Lower resolution chemical kinetics simulation (105 cells) discretization (10-100 zones)

Fluid mechanics sets the Combustion is very fast temperature distribution and therefore can be where autoignition occurs analyzed without

considering mixing or turbulence

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing

Head SAE Paper 2006-01-1363

Center line

Liner

Piston

Radial stratification imposed from centerline to liner ldquoUniformrdquo ldquoShallowrdquo and ldquoSteeprdquo

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

KIVA-MZ-MPI results compare well with

Pres

sure

(bar

)the results from a full integration of KIVA and Chemkin

90 350Phi = 026

Uniform (sim) 80 300Shallow (sim)

70 Steep (sim) 250

Rate of H

60 200

eat R

50 150

40 100

elease (JCA

D)

30 50

20 -20 -15 -10 -5 0 5 10 15

0 20

Crank Angle (Degree) φ=026

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors

8192 CPUs 8 TB Memory 122 Tflopssec

160 Tons 6 MW

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)

hellip

τ T

P

EGR

Ignition Delay Timeφ

Input Variables hellip

Input Hidden Output Layer Layers Layer

tk 1Ignition Condition I ( )t = intt (0) τ

dt = 1

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run

1 3

Is Ignition Criterion

met

No 4

Yes

Turn on global fuel conversion mechanism C8H18 + 85O2 rarr 8CO + 9H2O CO + frac12O2 rarr CO2

2 Calculate Ignition Integral in Each Cell

(0)

1( ) kt

t I t dt

τ = int

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

pres

sure

KPa

Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)

588 mm 523 mm 901 mm 0305 mm (02314 in) (0206 in) (0355 in) (00120 in)

8500

8000 Experimental Multi-zone

φ=026

7500 Neural network 024

022

7000 020

6500 018

102 mm (402 in) 6000 016

014 5500

φ=010 012 5000

4500

4000 -10 0 10

crank angle degrees Data from Sandia Livermore SAE 2003-01-0752

Scalloped piston iso-octane φ=01-026 CR=181 120 kPa intake

KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI

KIVA3V-MZ-MPI KIVA3V-ANN

φ=020 φ=020

KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI

KIVA3V-MZ-MPI KIVA3V-ANN

-4deg

-2deg

TDC

+4deg

+10 φ=020

-4deg

-2deg

TDC

+4deg

+10 φ=020

At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core

Kiva3v-MZ-MPI Kiva3v-ANN

φ=010 φ=010

Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model

Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI

KIVA3V-MZ-MPI KIVA3V-ANN

-4deg

-2deg

TDC

+4deg

+10 φ=020

-4deg

-2deg

TDC

+4deg

+10 φ=020

At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core

Kiva3v-MZ-MPI Kiva3v-ANN

φ=010 φ=010

Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model

Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core

Kiva3v-MZ-MPI Kiva3v-ANN

φ=010 φ=010

Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model

Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model

Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion

High resolution CFD Chemistry handled by multi-solver handles mixing zone detailed kinetics solveradvection and diffusion (10-100 zones)(~100k cells)

Solutions are mapped back and forth between solvers throughout the cycle

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency

bull Direct integration of KIVA and Chemkin

Years of computing time in single processor computer

bull Multi-zone KIVA-Chemkin Days of computing time in single processor computer

T

helliphellip

bull KIVA-Artificial neural networkP τ Ignition

Input Delayφ Time Hours of computing timeVariables

EGR

in single processor computerOutput Input Hidden

Layer Layers Layer

• Numerical Modeling of HCCI Combustion
• Homogeneous charge compression ignition (HCCI) combustion is an autoignition process controlled by chemical kinetics
• The physics of HCCI combustion can be well captured with a sequential fluid mechanics-chemical kinetics model
• Application of the multi-zone model Sandia engine running at low equivalence ratio with iso-octane
• Multi-zone model accurately predicts pressure traces for multiple equivalence ratios from ultra lean to mid load
• Multi-zone model makes good predictions of emissions over the full range of operation
• Model can tell us the location in the cylinder where different pollutants originate
• Much interest exists on Premixed Charge Compression Ignition (PCCI) engines for high load and improved combustion control
• Can we extend our sequential fluid mechanics-chemical kinetics model to model PCCI combustion
• We can try analyzing PCCI by doing a two-directional mapping from KIVA to CHEMKIN and from CHEMKIN back to KIVA
• We have applied KIVA-MZ-MPI to analyze partially stratified combustion with a linear fuel distribution at intake valve closing
• KIVA-MZ-MPI results compare well with the results from a full integration of KIVA and Chemkin
• KIVA-MZ-MPI gives significant reduction in computational time but still requires 1-2 days in 50-100 processors
• We have incorporated a neural network into KIVA3V for fast analysis of HCCI combustion and emissions (KIVA3V-ANN)
• KIVA3V-ANN greatly reduces computational intensity Firing cases take ~10 more time than a motored run
• Results agree well with experiments and multi-zone model over a wide range of conditions (geometry φ CR intake pressure)
• KIVA3V-ANN predicts similar spatial evolution of combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• KIVA3V-ANN predicts similar spatial evolution of CO during combustion as KIVA3V-MZ-MPI for ldquonormalrdquo HCCI
• At φ=010 KIVA3V-ANN predicts high CO emissions because reaction mechanism misses the oxidation of CO in the central core
• Summary HCCI is dominated by chemical kinetics and can therefore be well characterized by a computationally efficient segregated fluid mechanics-multi-zone chemical kinetics model
• Summary Our new fully integrated parallelized fluid mechanics-chemical kinetics code (KIVA3V-MZ-MPI) considers the effect of mixing and therefore applies to partially stratified combustion
• Summary We are developing accurate HCCI and PCCI analysis techniques with greatly improved computational efficiency