SlideShare a Scribd company logo
1 of 20
Download to read offline
Arrhenius.jl: A Differentiable Combustion Simulation Package
Weiqi Ji, Sili Deng
Department of Mechanical Engineering
Massachusetts Institute of Technology
May 25, 2021
1
12th U.S. National Combustion Meeting, Texas A&M University, Virtual
Paper ID: 2E06
2
❑ Complex kinetic models are not only expensive to
simulate but also challenging to develop
▪ More rate parameters to determine
❑ Experimental challenges
▪ Need larger datasets
▪ More important species to measure
▪ More experimental conditions to explore
❑ Computational challenges
▪ More parameters to optimize
▪ Evaluation is more expensive
Challenges for Developing Complex Kinetic Models
Adapted from Lu & Law, PECS, 2009
High-dimensional nonlinear optimization
3
❑ Evolutionary algorithms are usually employed for data-driven combustion modeling.
❑ Cost scales with number of parameters and number of data samples, not suitable for complex kinetic
models and large datasets.
Conventional Optimization without Differential Programming
Genetic
algorithms
Evolutionary
programming
Evolution
strategies
Particle
Swarm
Evolutionary algorithms
Simple models
< 100 parameters
Small datasets
Pros: easy implementation, multi-procs parallelization
Arrhenius.jl aims at optimizing complex models with hundreds and thousands of parameters,
utilizing large datasets from high throughput measurements.
4
❑ Efficient and robust optimization using
Stochastic Gradient Descent (SGD).
❑ Efficient and easy computation of gradients of
simulation results to model parameters.
Optimization with Differential Programming
Arrhenius.jl generalizes differential programing from deep learning to kinetic modeling.
LeCun et al., Nature 521 (2015) 436:444.
Differential programing refers to the optimization techniques for training deep neural networks.
5
▪ Datasets: Species profiles during propane
pyrolysis in shock tubes @Stanford.
▪ Optimization targets: 110 reactions related to
propane pyrolysis in USC Mech II (three
Arrhenius parameters for each reaction.
▪ Challenges: Highest dimension ever achieved
without parameter selection.
▪ Cost: Less than 100 lines of code; 4 hours on
a common PC (without HPC).
Example: Optimization of 330 Kinetic Parameters
Collaboration with Hanson’s group @Stanford
6
Framework of Arrhenius.jl
Kinetic
database
Thermodynamic
database
Transport
database
Mechanism
interpreter
Reaction
source term
Mixture properties
Applications
e.g., auto-ignition, jet stirred reactor, one-dimensional flames
Experimental data
solving
evaluation
differentiation
Neural Network
sub-models
Learning algorithms
Key Features
▪ Differentiable CHEMKIN
▪ Open source
▪ GPU support
▪ Augmented neural network
7
Differentiable Combustion Code
Differentiate computational blocks
Differentiate final solutions
𝜕F
𝜕𝜃 model parameters (e.g., A, b, Ea)
reaction source term, convection and diffusion terms
𝜕G
𝜕𝜃
species profiles, ignition delay, flame speed, extinction strain rate
model parameters (e.g., A, b, Ea), boundary/initial conditions, neural network submodels
convection diffusion reaction
8
Applications of Arrhenius.jl
Mechanism
Reduction
Uncertainty
Quantification
Inverse
Modeling
Model
Discovery
Data
Assimilation
Presented today
Preliminary results
9
❑ Conventional mechanism reduction keeps rate constants unchanged while removing pathways.
❑ Deep reduction: first over-reduce and then re-train to balance size and accuracy.
Mechanism Reduction
Over-reduced
mechanism
Re-train with
IDT data
Validate with
IDT / SL
P: 1 - 60 atm
T: 1100 - 2000 K
Phi: 0.5 - 1.8
Stochastic/mini-batch
Early-stopping
Weight decay
Fix key SL sensitive reactions
Compared to others
▪ Better fitness
▪ Better generalization
▪ IDT: Ignition Delay Time
▪ SL: Laminar Flame Speed
▪ Grad of IDT computed via SensBVP + auto-diff
SensBVP: Gururajan, and Egolfopoulos, CNF, 2019
10
Results: Deep Mechanism Reduction
▪ Performance: Deeply reduced models
work well for both IDT and SL.
▪ Advantage: Arrhenius.jl can optimize
hundreds of parameters, key to the
success of deep reduction.
▪ Potential: Further reduction is possible
for specific targets.
▪ Cost: Differential programing 1 hour on
a PC vs Evolution algorithm on an HPC
cluster
11
▪ Goal: Dimension reduction via PCA in gradient space
▪ Challenge: 4846 is the highest dimension ever
reached in combustion uncertainty quantification
▪ Insight: Observed 1D subspace for large HC,
suggesting low-cost UQ studies
▪ Cost: (GRI3.0) Arrhenius.jl ~ one minute on a laptop
vs Finite difference ~ one hour
Results: Uncertainty Quantification via Active Subspace
Active variable: project x into leading eigenvectors.
Arrhenius.jl tackles curse of dimensionality via efficiently identifying low-dimensional active subspaces.
12
Results: Autonomous Model Discovery
❑ Goal: Learn pyrolysis steps in HyChem without expert knowledges.
❑ Challenge: Augment NN into governing equations and optimize hundreds of NN weights.
CRNN (Chemical Reaction Neural Network):
An interpretable NN model with mass action
law and Arrhenius law encoded into structure.
Augment neural network sub model for autonomous scientific discovery.
Ji & Deng, JPC A, 2021
Talk: 3B01 @ Fire Research
13
Ongoing: Data Assimilation and Inverse Modeling
❑ Assimilate propane pyrolysis data from shock
tube, involving 330 parameters.
❑ Inference of unknown upstream with sparse
and noise measurements at downstream.
Collaboration with Hanson’s group @Stanford Submitted to ASME IMECE
Downstream
measurements
Upstream
parameterized
by neural
network
Inference
14
Open Source: Code, Discussions, Documents
❑ Code hosted on Github ❑ Online document
https://deng-mit.github.io/Arrhenius.jl/dev/
https://github.com/DENG-MIT/Arrhenius.jl
15
❑ Arrhenius.jl enables easy access to differentiable programing for optimizing complex combustion
kinetic models.
❑ Arrhenius.jl shows benefits of differential programing in many computational tasks.
Summary
Mechanism
Reduction
Uncertainty
Quantification
Inverse
Modeling
Model
Discovery
Data
Assimilation
▪ Deep reduction ▪ Active subspaces ▪ CRNN-HyChem
▪ Shock tube measurements ▪ Neural differential equations
16
Future Work
Functionality/Scalability
Multiphases/Multiphysics
▪ Solver for 1D flame (currently calling Cantera), extinction
▪ GPU support, matrix operation, batch-operation
▪ 2D/3D flame solver and gradient computing
▪ Scientific machine learning
▪ Super-critical, real gas states
▪ Condensed phases: soot, flame synthesis, energetic materials, fire
▪ Electrochemistry, catalyst, plasma
▪ Optics: differentiable ray-tracing, laser spectroscopy
We welcome collaborators,
contributors and users join the
journey in all means
Arrhenius.jl enables fast incorporation of new physics via its readability, efficiency, multi-dispatch.
Publications
Applications are also open-source
https://github.com/DENG-MIT
▪ DeepReduction
▪ ArrheniusActiveSubspace
▪ Arrhenius_Flame_1D
▪ NN-ShockTube
▪ …
18
❑ Developers
▪ @jiweiqi (MIT)
▪ @TJP-Karpowski (TU Darmstadt)
▪ @SuXY15 (Tsinghua)
▪ @RSuryaNarayan (NIT, Tiruchirappalli)
❑ Lead Developer of SciML
▪ @ChrisRackauckas
▪ @YingboMa
❑ Lead Developer of RMS.jl
▪ @mjohnson541 (MIT)
❑ Cantera’s open-source community
❑ Vyaas Gururajan (ANL): SensBVP
❑ Ji-Woong Park (ANL): deep reduction
Acknowledgements
Backup Slides
20
SensBVP for Ignition Delay Time
𝑑𝑦
𝑑𝑡′ = 𝜏𝑓 𝑦, 𝜏𝑡′ , 𝑡 = 𝜏𝑡′
B.C. 𝑡′
= 0, y = 𝑦0
𝑡′ = 1, y = 𝑦𝜏 (T = Tign)
▪ Initial Value Problem
(IVP)
▪ Boundary Value Problem
(BVP)
𝐹(𝛼) = 0
𝜕𝐹
𝜕𝑦
𝜕𝑦
𝜕𝛼
+
𝜕𝐹
𝜕𝛼
= 0
▪ SensBVP
𝑑𝑦
𝑑𝑡
= 𝑓 𝑦, 𝑡
Auto-Differentiation
𝜕𝐹
𝜕𝑦
,
𝜕𝐹
𝜕𝛼
Gururajan, Vyaas, and Fokion N. Egolfopoulos. "Direct sensitivity analysis for ignition delay
times." Combustion and Flame 209 (2019): 478-480.

More Related Content

What's hot

TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...
TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...
TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...Journal For Research
 
Mid year research report
Mid year research reportMid year research report
Mid year research reportGavaga
 
Task scheduling Survey in Cloud Computing
Task scheduling Survey in Cloud ComputingTask scheduling Survey in Cloud Computing
Task scheduling Survey in Cloud ComputingRamandeep Kaur
 
Cost Efficient PageRank Computation using GPU : NOTES
Cost Efficient PageRank Computation using GPU : NOTESCost Efficient PageRank Computation using GPU : NOTES
Cost Efficient PageRank Computation using GPU : NOTESSubhajit Sahu
 
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...IEEEGLOBALSOFTSTUDENTPROJECTS
 
Pdcs2010 balman-presentation
Pdcs2010 balman-presentationPdcs2010 balman-presentation
Pdcs2010 balman-presentationbalmanme
 
DuraMat Data Analytics
DuraMat Data AnalyticsDuraMat Data Analytics
DuraMat Data AnalyticsAnubhav Jain
 
Sky Arrays - ArrayDB in action for Sky View Factor Computation
Sky Arrays - ArrayDB in action for Sky View Factor ComputationSky Arrays - ArrayDB in action for Sky View Factor Computation
Sky Arrays - ArrayDB in action for Sky View Factor ComputationEUDAT
 
ThesisDefensePresentation_KyleIngersoll
ThesisDefensePresentation_KyleIngersollThesisDefensePresentation_KyleIngersoll
ThesisDefensePresentation_KyleIngersollKyle Ingersoll
 
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMJOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMmailjkb
 
Cloud Computing
Cloud ComputingCloud Computing
Cloud Computingbutest
 
Scalable Parallel Computing on Clouds
Scalable Parallel Computing on CloudsScalable Parallel Computing on Clouds
Scalable Parallel Computing on CloudsThilina Gunarathne
 
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...Ptidej Team
 
Recent developments in HPX and Octo-Tiger
Recent developments in HPX and Octo-TigerRecent developments in HPX and Octo-Tiger
Recent developments in HPX and Octo-TigerPatrick Diehl
 

What's hot (18)

TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...
TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...
TASK SCHEDULING USING AMALGAMATION OF MET HEURISTICS SWARM OPTIMIZATION ALGOR...
 
19 characterizing pv modules using microinverter data final
19 characterizing pv modules using microinverter data   final19 characterizing pv modules using microinverter data   final
19 characterizing pv modules using microinverter data final
 
Mid year research report
Mid year research reportMid year research report
Mid year research report
 
CLOUD BIOINFORMATICS Part1
 CLOUD BIOINFORMATICS Part1 CLOUD BIOINFORMATICS Part1
CLOUD BIOINFORMATICS Part1
 
Task scheduling Survey in Cloud Computing
Task scheduling Survey in Cloud ComputingTask scheduling Survey in Cloud Computing
Task scheduling Survey in Cloud Computing
 
Cost Efficient PageRank Computation using GPU : NOTES
Cost Efficient PageRank Computation using GPU : NOTESCost Efficient PageRank Computation using GPU : NOTES
Cost Efficient PageRank Computation using GPU : NOTES
 
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...
IEEE 2014 JAVA CLOUD COMPUTING PROJECTS Adaptive algorithm for minimizing clo...
 
Pdcs2010 balman-presentation
Pdcs2010 balman-presentationPdcs2010 balman-presentation
Pdcs2010 balman-presentation
 
DuraMat Data Analytics
DuraMat Data AnalyticsDuraMat Data Analytics
DuraMat Data Analytics
 
Sky Arrays - ArrayDB in action for Sky View Factor Computation
Sky Arrays - ArrayDB in action for Sky View Factor ComputationSky Arrays - ArrayDB in action for Sky View Factor Computation
Sky Arrays - ArrayDB in action for Sky View Factor Computation
 
ThesisDefensePresentation_KyleIngersoll
ThesisDefensePresentation_KyleIngersollThesisDefensePresentation_KyleIngersoll
ThesisDefensePresentation_KyleIngersoll
 
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHMJOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
JOB SCHEDULING USING ANT COLONY OPTIMIZATION ALGORITHM
 
Thesis
ThesisThesis
Thesis
 
Cloud Computing
Cloud ComputingCloud Computing
Cloud Computing
 
Svm on cloud (presntation)
Svm on cloud  (presntation)Svm on cloud  (presntation)
Svm on cloud (presntation)
 
Scalable Parallel Computing on Clouds
Scalable Parallel Computing on CloudsScalable Parallel Computing on Clouds
Scalable Parallel Computing on Clouds
 
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...
140829+an+empirical+study+of+the+impact+of+cloud+patterns+on+quality+of+servi...
 
Recent developments in HPX and Octo-Tiger
Recent developments in HPX and Octo-TigerRecent developments in HPX and Octo-Tiger
Recent developments in HPX and Octo-Tiger
 

Similar to Arrhenius.jl: A Differentiable Combustion Simulation Package

Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...
Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...
Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...eArtius, Inc.
 
Accelerators at ORNL - Application Readiness, Early Science, and Industry Impact
Accelerators at ORNL - Application Readiness, Early Science, and Industry ImpactAccelerators at ORNL - Application Readiness, Early Science, and Industry Impact
Accelerators at ORNL - Application Readiness, Early Science, and Industry Impactinside-BigData.com
 
Software tools, crystal descriptors, and machine learning applied to material...
Software tools, crystal descriptors, and machine learning applied to material...Software tools, crystal descriptors, and machine learning applied to material...
Software tools, crystal descriptors, and machine learning applied to material...Anubhav Jain
 
Accelerating S3D A GPGPU Case Study
Accelerating S3D  A GPGPU Case StudyAccelerating S3D  A GPGPU Case Study
Accelerating S3D A GPGPU Case StudyMartha Brown
 
Towards Exascale Engine Simulations with NEK5000
Towards Exascale Engine Simulations with NEK5000Towards Exascale Engine Simulations with NEK5000
Towards Exascale Engine Simulations with NEK5000inside-BigData.com
 
byteLAKE's expertise across NVIDIA architectures and configurations
byteLAKE's expertise across NVIDIA architectures and configurationsbyteLAKE's expertise across NVIDIA architectures and configurations
byteLAKE's expertise across NVIDIA architectures and configurationsbyteLAKE
 
2007 Oral Preliminary Defense
2007 Oral Preliminary Defense2007 Oral Preliminary Defense
2007 Oral Preliminary DefenseJon Ernstberger
 
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
AI optimizing HPC simulations (presentation from  6th EULAG Workshop)AI optimizing HPC simulations (presentation from  6th EULAG Workshop)
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)byteLAKE
 
CloudLightning and the OPM-based Use Case
CloudLightning and the OPM-based Use CaseCloudLightning and the OPM-based Use Case
CloudLightning and the OPM-based Use CaseCloudLightning
 
Early Application experiences on Summit
Early Application experiences on Summit Early Application experiences on Summit
Early Application experiences on Summit Ganesan Narayanasamy
 
The U.S. Exascale Computing Project: Status and Plans
The U.S. Exascale Computing Project: Status and PlansThe U.S. Exascale Computing Project: Status and Plans
The U.S. Exascale Computing Project: Status and Plansinside-BigData.com
 
LSI Seminar on Marina Zapater's PhD Thesis
LSI Seminar on Marina Zapater's PhD ThesisLSI Seminar on Marina Zapater's PhD Thesis
LSI Seminar on Marina Zapater's PhD ThesisGreenLSI Team, LSI, UPM
 
QuantumChemistry500
QuantumChemistry500QuantumChemistry500
QuantumChemistry500Maho Nakata
 
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...IJERD Editor
 
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...IJERD Editor
 
Gray-Box Models for Performance Assessment of Spark Applications
Gray-Box Models for Performance Assessment of Spark ApplicationsGray-Box Models for Performance Assessment of Spark Applications
Gray-Box Models for Performance Assessment of Spark ApplicationsATMOSPHERE .
 

Similar to Arrhenius.jl: A Differentiable Combustion Simulation Package (20)

ECP Application Development
ECP Application DevelopmentECP Application Development
ECP Application Development
 
Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...
Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...
Multi-Objective Optimization of Solar Cells Thermal Uniformity Using Combined...
 
Presentation
PresentationPresentation
Presentation
 
Accelerators at ORNL - Application Readiness, Early Science, and Industry Impact
Accelerators at ORNL - Application Readiness, Early Science, and Industry ImpactAccelerators at ORNL - Application Readiness, Early Science, and Industry Impact
Accelerators at ORNL - Application Readiness, Early Science, and Industry Impact
 
Software tools, crystal descriptors, and machine learning applied to material...
Software tools, crystal descriptors, and machine learning applied to material...Software tools, crystal descriptors, and machine learning applied to material...
Software tools, crystal descriptors, and machine learning applied to material...
 
Fahroo - Computational Mathematics - Spring Review 2012
Fahroo - Computational Mathematics - Spring Review 2012 Fahroo - Computational Mathematics - Spring Review 2012
Fahroo - Computational Mathematics - Spring Review 2012
 
Accelerating S3D A GPGPU Case Study
Accelerating S3D  A GPGPU Case StudyAccelerating S3D  A GPGPU Case Study
Accelerating S3D A GPGPU Case Study
 
Towards Exascale Engine Simulations with NEK5000
Towards Exascale Engine Simulations with NEK5000Towards Exascale Engine Simulations with NEK5000
Towards Exascale Engine Simulations with NEK5000
 
byteLAKE's expertise across NVIDIA architectures and configurations
byteLAKE's expertise across NVIDIA architectures and configurationsbyteLAKE's expertise across NVIDIA architectures and configurations
byteLAKE's expertise across NVIDIA architectures and configurations
 
2007 Oral Preliminary Defense
2007 Oral Preliminary Defense2007 Oral Preliminary Defense
2007 Oral Preliminary Defense
 
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
AI optimizing HPC simulations (presentation from  6th EULAG Workshop)AI optimizing HPC simulations (presentation from  6th EULAG Workshop)
AI optimizing HPC simulations (presentation from 6th EULAG Workshop)
 
CloudLightning and the OPM-based Use Case
CloudLightning and the OPM-based Use CaseCloudLightning and the OPM-based Use Case
CloudLightning and the OPM-based Use Case
 
Early Application experiences on Summit
Early Application experiences on Summit Early Application experiences on Summit
Early Application experiences on Summit
 
The U.S. Exascale Computing Project: Status and Plans
The U.S. Exascale Computing Project: Status and PlansThe U.S. Exascale Computing Project: Status and Plans
The U.S. Exascale Computing Project: Status and Plans
 
LSI Seminar on Marina Zapater's PhD Thesis
LSI Seminar on Marina Zapater's PhD ThesisLSI Seminar on Marina Zapater's PhD Thesis
LSI Seminar on Marina Zapater's PhD Thesis
 
QuantumChemistry500
QuantumChemistry500QuantumChemistry500
QuantumChemistry500
 
Fahroo - Computational Mathematics - Spring Review 2013
Fahroo - Computational Mathematics - Spring Review 2013Fahroo - Computational Mathematics - Spring Review 2013
Fahroo - Computational Mathematics - Spring Review 2013
 
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
 
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
Neuro-Genetic Optimization of LDO-fired Rotary Furnace Parameters for the Pro...
 
Gray-Box Models for Performance Assessment of Spark Applications
Gray-Box Models for Performance Assessment of Spark ApplicationsGray-Box Models for Performance Assessment of Spark Applications
Gray-Box Models for Performance Assessment of Spark Applications
 

Recently uploaded

Biopesticide (2).pptx .This slides helps to know the different types of biop...
Biopesticide (2).pptx  .This slides helps to know the different types of biop...Biopesticide (2).pptx  .This slides helps to know the different types of biop...
Biopesticide (2).pptx .This slides helps to know the different types of biop...RohitNehra6
 
Scheme-of-Work-Science-Stage-4 cambridge science.docx
Scheme-of-Work-Science-Stage-4 cambridge science.docxScheme-of-Work-Science-Stage-4 cambridge science.docx
Scheme-of-Work-Science-Stage-4 cambridge science.docxyaramohamed343013
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...Sérgio Sacani
 
Genomic DNA And Complementary DNA Libraries construction.
Genomic DNA And Complementary DNA Libraries construction.Genomic DNA And Complementary DNA Libraries construction.
Genomic DNA And Complementary DNA Libraries construction.k64182334
 
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...Sérgio Sacani
 
Boyles law module in the grade 10 science
Boyles law module in the grade 10 scienceBoyles law module in the grade 10 science
Boyles law module in the grade 10 sciencefloriejanemacaya1
 
Isotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoIsotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoSérgio Sacani
 
Physiochemical properties of nanomaterials and its nanotoxicity.pptx
Physiochemical properties of nanomaterials and its nanotoxicity.pptxPhysiochemical properties of nanomaterials and its nanotoxicity.pptx
Physiochemical properties of nanomaterials and its nanotoxicity.pptxAArockiyaNisha
 
zoogeography of pakistan.pptx fauna of Pakistan
zoogeography of pakistan.pptx fauna of Pakistanzoogeography of pakistan.pptx fauna of Pakistan
zoogeography of pakistan.pptx fauna of Pakistanzohaibmir069
 
Recombinant DNA technology( Transgenic plant and animal)
Recombinant DNA technology( Transgenic plant and animal)Recombinant DNA technology( Transgenic plant and animal)
Recombinant DNA technology( Transgenic plant and animal)DHURKADEVIBASKAR
 
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.PraveenaKalaiselvan1
 
Module 4: Mendelian Genetics and Punnett Square
Module 4:  Mendelian Genetics and Punnett SquareModule 4:  Mendelian Genetics and Punnett Square
Module 4: Mendelian Genetics and Punnett SquareIsiahStephanRadaza
 
Orientation, design and principles of polyhouse
Orientation, design and principles of polyhouseOrientation, design and principles of polyhouse
Orientation, design and principles of polyhousejana861314
 
Unlocking the Potential: Deep dive into ocean of Ceramic Magnets.pptx
Unlocking  the Potential: Deep dive into ocean of Ceramic Magnets.pptxUnlocking  the Potential: Deep dive into ocean of Ceramic Magnets.pptx
Unlocking the Potential: Deep dive into ocean of Ceramic Magnets.pptxanandsmhk
 
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfAnalytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfSwapnil Therkar
 
Work, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE PhysicsWork, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE Physicsvishikhakeshava1
 
Artificial Intelligence In Microbiology by Dr. Prince C P
Artificial Intelligence In Microbiology by Dr. Prince C PArtificial Intelligence In Microbiology by Dr. Prince C P
Artificial Intelligence In Microbiology by Dr. Prince C PPRINCE C P
 
The Black hole shadow in Modified Gravity
The Black hole shadow in Modified GravityThe Black hole shadow in Modified Gravity
The Black hole shadow in Modified GravitySubhadipsau21168
 
Disentangling the origin of chemical differences using GHOST
Disentangling the origin of chemical differences using GHOSTDisentangling the origin of chemical differences using GHOST
Disentangling the origin of chemical differences using GHOSTSérgio Sacani
 

Recently uploaded (20)

Biopesticide (2).pptx .This slides helps to know the different types of biop...
Biopesticide (2).pptx  .This slides helps to know the different types of biop...Biopesticide (2).pptx  .This slides helps to know the different types of biop...
Biopesticide (2).pptx .This slides helps to know the different types of biop...
 
Scheme-of-Work-Science-Stage-4 cambridge science.docx
Scheme-of-Work-Science-Stage-4 cambridge science.docxScheme-of-Work-Science-Stage-4 cambridge science.docx
Scheme-of-Work-Science-Stage-4 cambridge science.docx
 
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
 
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
PossibleEoarcheanRecordsoftheGeomagneticFieldPreservedintheIsuaSupracrustalBe...
 
Genomic DNA And Complementary DNA Libraries construction.
Genomic DNA And Complementary DNA Libraries construction.Genomic DNA And Complementary DNA Libraries construction.
Genomic DNA And Complementary DNA Libraries construction.
 
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
Discovery of an Accretion Streamer and a Slow Wide-angle Outflow around FUOri...
 
Boyles law module in the grade 10 science
Boyles law module in the grade 10 scienceBoyles law module in the grade 10 science
Boyles law module in the grade 10 science
 
Isotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on IoIsotopic evidence of long-lived volcanism on Io
Isotopic evidence of long-lived volcanism on Io
 
Physiochemical properties of nanomaterials and its nanotoxicity.pptx
Physiochemical properties of nanomaterials and its nanotoxicity.pptxPhysiochemical properties of nanomaterials and its nanotoxicity.pptx
Physiochemical properties of nanomaterials and its nanotoxicity.pptx
 
zoogeography of pakistan.pptx fauna of Pakistan
zoogeography of pakistan.pptx fauna of Pakistanzoogeography of pakistan.pptx fauna of Pakistan
zoogeography of pakistan.pptx fauna of Pakistan
 
Recombinant DNA technology( Transgenic plant and animal)
Recombinant DNA technology( Transgenic plant and animal)Recombinant DNA technology( Transgenic plant and animal)
Recombinant DNA technology( Transgenic plant and animal)
 
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.
BIOETHICS IN RECOMBINANT DNA TECHNOLOGY.
 
Module 4: Mendelian Genetics and Punnett Square
Module 4:  Mendelian Genetics and Punnett SquareModule 4:  Mendelian Genetics and Punnett Square
Module 4: Mendelian Genetics and Punnett Square
 
Orientation, design and principles of polyhouse
Orientation, design and principles of polyhouseOrientation, design and principles of polyhouse
Orientation, design and principles of polyhouse
 
Unlocking the Potential: Deep dive into ocean of Ceramic Magnets.pptx
Unlocking  the Potential: Deep dive into ocean of Ceramic Magnets.pptxUnlocking  the Potential: Deep dive into ocean of Ceramic Magnets.pptx
Unlocking the Potential: Deep dive into ocean of Ceramic Magnets.pptx
 
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdfAnalytical Profile of Coleus Forskohlii | Forskolin .pdf
Analytical Profile of Coleus Forskohlii | Forskolin .pdf
 
Work, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE PhysicsWork, Energy and Power for class 10 ICSE Physics
Work, Energy and Power for class 10 ICSE Physics
 
Artificial Intelligence In Microbiology by Dr. Prince C P
Artificial Intelligence In Microbiology by Dr. Prince C PArtificial Intelligence In Microbiology by Dr. Prince C P
Artificial Intelligence In Microbiology by Dr. Prince C P
 
The Black hole shadow in Modified Gravity
The Black hole shadow in Modified GravityThe Black hole shadow in Modified Gravity
The Black hole shadow in Modified Gravity
 
Disentangling the origin of chemical differences using GHOST
Disentangling the origin of chemical differences using GHOSTDisentangling the origin of chemical differences using GHOST
Disentangling the origin of chemical differences using GHOST
 

Arrhenius.jl: A Differentiable Combustion Simulation Package

  • 1. Arrhenius.jl: A Differentiable Combustion Simulation Package Weiqi Ji, Sili Deng Department of Mechanical Engineering Massachusetts Institute of Technology May 25, 2021 1 12th U.S. National Combustion Meeting, Texas A&M University, Virtual Paper ID: 2E06
  • 2. 2 ❑ Complex kinetic models are not only expensive to simulate but also challenging to develop ▪ More rate parameters to determine ❑ Experimental challenges ▪ Need larger datasets ▪ More important species to measure ▪ More experimental conditions to explore ❑ Computational challenges ▪ More parameters to optimize ▪ Evaluation is more expensive Challenges for Developing Complex Kinetic Models Adapted from Lu & Law, PECS, 2009 High-dimensional nonlinear optimization
  • 3. 3 ❑ Evolutionary algorithms are usually employed for data-driven combustion modeling. ❑ Cost scales with number of parameters and number of data samples, not suitable for complex kinetic models and large datasets. Conventional Optimization without Differential Programming Genetic algorithms Evolutionary programming Evolution strategies Particle Swarm Evolutionary algorithms Simple models < 100 parameters Small datasets Pros: easy implementation, multi-procs parallelization Arrhenius.jl aims at optimizing complex models with hundreds and thousands of parameters, utilizing large datasets from high throughput measurements.
  • 4. 4 ❑ Efficient and robust optimization using Stochastic Gradient Descent (SGD). ❑ Efficient and easy computation of gradients of simulation results to model parameters. Optimization with Differential Programming Arrhenius.jl generalizes differential programing from deep learning to kinetic modeling. LeCun et al., Nature 521 (2015) 436:444. Differential programing refers to the optimization techniques for training deep neural networks.
  • 5. 5 ▪ Datasets: Species profiles during propane pyrolysis in shock tubes @Stanford. ▪ Optimization targets: 110 reactions related to propane pyrolysis in USC Mech II (three Arrhenius parameters for each reaction. ▪ Challenges: Highest dimension ever achieved without parameter selection. ▪ Cost: Less than 100 lines of code; 4 hours on a common PC (without HPC). Example: Optimization of 330 Kinetic Parameters Collaboration with Hanson’s group @Stanford
  • 6. 6 Framework of Arrhenius.jl Kinetic database Thermodynamic database Transport database Mechanism interpreter Reaction source term Mixture properties Applications e.g., auto-ignition, jet stirred reactor, one-dimensional flames Experimental data solving evaluation differentiation Neural Network sub-models Learning algorithms Key Features ▪ Differentiable CHEMKIN ▪ Open source ▪ GPU support ▪ Augmented neural network
  • 7. 7 Differentiable Combustion Code Differentiate computational blocks Differentiate final solutions 𝜕F 𝜕𝜃 model parameters (e.g., A, b, Ea) reaction source term, convection and diffusion terms 𝜕G 𝜕𝜃 species profiles, ignition delay, flame speed, extinction strain rate model parameters (e.g., A, b, Ea), boundary/initial conditions, neural network submodels convection diffusion reaction
  • 9. 9 ❑ Conventional mechanism reduction keeps rate constants unchanged while removing pathways. ❑ Deep reduction: first over-reduce and then re-train to balance size and accuracy. Mechanism Reduction Over-reduced mechanism Re-train with IDT data Validate with IDT / SL P: 1 - 60 atm T: 1100 - 2000 K Phi: 0.5 - 1.8 Stochastic/mini-batch Early-stopping Weight decay Fix key SL sensitive reactions Compared to others ▪ Better fitness ▪ Better generalization ▪ IDT: Ignition Delay Time ▪ SL: Laminar Flame Speed ▪ Grad of IDT computed via SensBVP + auto-diff SensBVP: Gururajan, and Egolfopoulos, CNF, 2019
  • 10. 10 Results: Deep Mechanism Reduction ▪ Performance: Deeply reduced models work well for both IDT and SL. ▪ Advantage: Arrhenius.jl can optimize hundreds of parameters, key to the success of deep reduction. ▪ Potential: Further reduction is possible for specific targets. ▪ Cost: Differential programing 1 hour on a PC vs Evolution algorithm on an HPC cluster
  • 11. 11 ▪ Goal: Dimension reduction via PCA in gradient space ▪ Challenge: 4846 is the highest dimension ever reached in combustion uncertainty quantification ▪ Insight: Observed 1D subspace for large HC, suggesting low-cost UQ studies ▪ Cost: (GRI3.0) Arrhenius.jl ~ one minute on a laptop vs Finite difference ~ one hour Results: Uncertainty Quantification via Active Subspace Active variable: project x into leading eigenvectors. Arrhenius.jl tackles curse of dimensionality via efficiently identifying low-dimensional active subspaces.
  • 12. 12 Results: Autonomous Model Discovery ❑ Goal: Learn pyrolysis steps in HyChem without expert knowledges. ❑ Challenge: Augment NN into governing equations and optimize hundreds of NN weights. CRNN (Chemical Reaction Neural Network): An interpretable NN model with mass action law and Arrhenius law encoded into structure. Augment neural network sub model for autonomous scientific discovery. Ji & Deng, JPC A, 2021 Talk: 3B01 @ Fire Research
  • 13. 13 Ongoing: Data Assimilation and Inverse Modeling ❑ Assimilate propane pyrolysis data from shock tube, involving 330 parameters. ❑ Inference of unknown upstream with sparse and noise measurements at downstream. Collaboration with Hanson’s group @Stanford Submitted to ASME IMECE Downstream measurements Upstream parameterized by neural network Inference
  • 14. 14 Open Source: Code, Discussions, Documents ❑ Code hosted on Github ❑ Online document https://deng-mit.github.io/Arrhenius.jl/dev/ https://github.com/DENG-MIT/Arrhenius.jl
  • 15. 15 ❑ Arrhenius.jl enables easy access to differentiable programing for optimizing complex combustion kinetic models. ❑ Arrhenius.jl shows benefits of differential programing in many computational tasks. Summary Mechanism Reduction Uncertainty Quantification Inverse Modeling Model Discovery Data Assimilation ▪ Deep reduction ▪ Active subspaces ▪ CRNN-HyChem ▪ Shock tube measurements ▪ Neural differential equations
  • 16. 16 Future Work Functionality/Scalability Multiphases/Multiphysics ▪ Solver for 1D flame (currently calling Cantera), extinction ▪ GPU support, matrix operation, batch-operation ▪ 2D/3D flame solver and gradient computing ▪ Scientific machine learning ▪ Super-critical, real gas states ▪ Condensed phases: soot, flame synthesis, energetic materials, fire ▪ Electrochemistry, catalyst, plasma ▪ Optics: differentiable ray-tracing, laser spectroscopy We welcome collaborators, contributors and users join the journey in all means Arrhenius.jl enables fast incorporation of new physics via its readability, efficiency, multi-dispatch.
  • 17. Publications Applications are also open-source https://github.com/DENG-MIT ▪ DeepReduction ▪ ArrheniusActiveSubspace ▪ Arrhenius_Flame_1D ▪ NN-ShockTube ▪ …
  • 18. 18 ❑ Developers ▪ @jiweiqi (MIT) ▪ @TJP-Karpowski (TU Darmstadt) ▪ @SuXY15 (Tsinghua) ▪ @RSuryaNarayan (NIT, Tiruchirappalli) ❑ Lead Developer of SciML ▪ @ChrisRackauckas ▪ @YingboMa ❑ Lead Developer of RMS.jl ▪ @mjohnson541 (MIT) ❑ Cantera’s open-source community ❑ Vyaas Gururajan (ANL): SensBVP ❑ Ji-Woong Park (ANL): deep reduction Acknowledgements
  • 20. 20 SensBVP for Ignition Delay Time 𝑑𝑦 𝑑𝑡′ = 𝜏𝑓 𝑦, 𝜏𝑡′ , 𝑡 = 𝜏𝑡′ B.C. 𝑡′ = 0, y = 𝑦0 𝑡′ = 1, y = 𝑦𝜏 (T = Tign) ▪ Initial Value Problem (IVP) ▪ Boundary Value Problem (BVP) 𝐹(𝛼) = 0 𝜕𝐹 𝜕𝑦 𝜕𝑦 𝜕𝛼 + 𝜕𝐹 𝜕𝛼 = 0 ▪ SensBVP 𝑑𝑦 𝑑𝑡 = 𝑓 𝑦, 𝑡 Auto-Differentiation 𝜕𝐹 𝜕𝑦 , 𝜕𝐹 𝜕𝛼 Gururajan, Vyaas, and Fokion N. Egolfopoulos. "Direct sensitivity analysis for ignition delay times." Combustion and Flame 209 (2019): 478-480.