ChemKANs for combustion chemistry modeling and acceleration

Area of Science:

• Combustion Chemistry
• Chemical Kinetics
• Machine Learning

Background:

• Chemical kinetic model inference and application in combustion face challenges due to large ODE systems and disparate time scales.
• Machine learning (ML) offers potential for streamlining these models, but nonlinearity, stiffness, and noisy data pose difficulties.

Purpose of the Study:

• Introduce ChemKANs, a novel neural network framework for combustion chemistry model inference and simulation acceleration.
• Enhance the generic Kolmogorov-Arnold network ordinary differential equations (KAN-ODEs) with chemical kinetic and thermodynamic principles.

Main Methods:

• Developed ChemKANs by augmenting KAN-ODEs with chemistry-specific knowledge of information flow.
• Benchmarked ChemKANs for model inference robustness against sparse, noisy data and large parameterizations.
• Evaluated ChemKANs for simulation acceleration of hydrogen combustion chemistry.

Main Results:

• ChemKANs demonstrated strong inductive bias, streamlined training, and higher accuracy compared to benchmarks.
• Exhibited no overfitting or degradation with up to 15% data noise and large parameterizations, showing resilience.
• A parameter-lean ChemKAN (344 parameters) accurately represented hydrogen combustion, achieving 2x acceleration.

Conclusions:

• ChemKANs offer a robust, expressive, and efficient approach for chemical kinetic model inference in combustion.
• The framework shows significant potential for accelerating simulations in combustion physics and chemical kinetics.
• ChemKANs are resilient to common deep learning failure modes, making them reliable for complex scientific applications.