Chemical Kinetics Reduction Using a Self-Evolving Neural Network Framework

The development of computationally efficient yet accurate chemical kinetic models is critical for enabling high-fidelity simulations of turbulent reacting flows. On the other hand, accurate representation of chemical characteristics requires the use of detailed reaction mechanisms, which comprise intricate networks of species and reactions, imposing substantial computational costs that hinder their practical use in large-scale simulations. To address this, we introduce a self-evolving neural network (SENN) framework for automated and systematic reduction of large kinetics, to accelerate the computation of chemical reactions in simulations of turbulent reacting flow. The SENN begins as a fully connected graphical network that links input species, reaction neurons, and net species production rates. Through iterative training episodes, the framework dynamically adapts by leveraging Hebbian learning to robustly reinforce pathways encoding the most influential reaction steps, while graph-guided pruning systematically excises weak or redundant connections, driving the network toward an optimally sparse and physically meaningful architecture. The reduction process advances in two stages: stage I applies topology-guided pruning and Hebbian learning to evolve the network toward the ground-truth kinetic mechanism; and stage II integrates cumulative rate-of-progress (ROP) and sensitivity analyses to identify and remove reaction neurons that contribute negligibly to overall kinetics and critical combustion metrics (e.g., ignition delay, laminar flame speed). Trained across a broad thermodynamic space, the final sparse and interpretable network retains essential chemical characteristics such as flame speed and ignition delay while dramatically reducing model complexity. In this presentation, we demonstrate the SENN framework on a well-known hydrogen-oxygen chemical reaction mechanism.