20.05.2025 - 13:05:11

Predicting Chemical Source Terms with Neural Networks

In reactive CFD simulations, evaluating chemical source terms with traditional ODE solvers is computationally expensive, especially for detailed mechanisms. Neural networks offer a faster surrogate, but their raw outputs can violate mass and element conservation. Our work introduces a lightweight, model-agnostic post-processing layer that applies a stoichiometry-based projection into the valid solution space, minimally adjusting each species by its relative importance. The proposed method corrects conservation violations without altering the network architecture or training process, delivering physically valid outputs even in mechanisms involving both major and minor species and is applicable to any system requiring mass or element conservation.