11.03.2026 - 10:47:16

New method dramatically reduces cost of physics-consistent neural networks for combustion simulations

M.Sc. Maximilian Schäfer has developed a new method to enforce elemental and mass conservation in neural networks predicting chemical source terms. The approach applies a weighted projection as a final correction layer, ensuring strict physical consistency without modifying or retraining the neural network. Compared to current state-of-the-art conservation methods, the new technique reduces computational overhead by roughly an order of magnitude while maintaining high predictive accuracy. The work has been accepted for presentation at the 41st International Symposium on Combustion, which will take place this summer in Japan.