Accelerating Combustion Modeling with DeepONet: A Machine Learning Approach for Stiff Chemical Kinetics
Stiff chemical kinetics make high-fidelity combustion simulations slow and often impractical, especially for fast or real-time predictions. We developed a machine learning approach using Deep Operator Networks (DeepONet) to overcome this bottleneck. DeepONet has two parts: a branch network that processes input data (like temperature or species curves), and a trunk network that handles where or when we want predictions. We tested two versions: one with ResNet in both parts, and another (DeepOKAN) with ResNet in the branch and KAN in the trunk. To handle long times, we divided the timeline into chunks and predicted each segment assuming known initial conditions. We also tested adaptive loss functions and two training strategies: standard (all-at-once) and two-step (branch/trunk separately). Adaptive loss improved accuracy, and two-step training performed slightly better. We demonstrated the method on syngas combustion and achieved high accuracy. We will develop a DeepONet-CK library for multiple fuels, targeting CFD tools like DTRA’s Ravel for decision support (IMEA) and HPC workflows.
IMPACT
Accomplishment: Developed a Deep Operator Network (DeepONet)–based machine learning surrogate to replace stiff chemical kinetics solvers in combustion simulations; Result: Achieved up to orders of magnitudes faster predictions with high accuracy for reacting flows – Enables near real-time simulation and decision-making for warfighter-relevant energetic systems, improving mission planning, system design, and response times in dynamic operational environments.
PRESENTER
Wei, Liang
wei@kcse.com
818-844-1995
Karagozian & Case, Inc.
CO-AUTHOR(S)
Abraham, Joseph
abraham@kcse.com
Reynolds, Adam
reynolds@kcse.com
Karniadakis, George
George_Karniadakis@brown.edu
Nath, Kamaljyoti
kamaljyoti_nath@brown.edu
Pandey, additi
additi_pandey@brown.edu
Babaee, Hessam
h.babaee@pitt.edu
Susi, Bryan
bsusi@ara.com
CATEGORY
AI/ML for HPC
SECONDARY CATEGORY
Comp Fluid Dynamics
SYSTEM(S) USED
Warhawk