On-the-fly accuracy evaluation of artificial neural networks and hybrid method to improve the robustness of neural network accelerated chemistry solving

Abstract

A chemistry acceleration strategy based on the coupling of artificial neural networks (ANNs) and direct integration (DI) is proposed and evaluated in the context of turbulent combustion. The main novelty of this study is its focus on ANNs robustness assessment. A hybrid DI/ANN strategy is proposed, which allows for a direct control of the prediction errors. This control is achieved by evaluating the ANN prediction error after each inference. To this end, a simple yet novel criterion based on mass conservation is proposed and compared to a criterion based on the distance between the inferred state and the training database, as done previously in the literature in the context of on-the-fly learning. A two-dimensional turbulent premixed H2 ignition case is used to assess the performance of the strategy and challenge the two criteria. An a priori study demonstrates that the state space-based criterion cannot correctly describe the ANN error, while the mass conservation-based one provides a good match with the ANN error. An a posteriori evaluation, involving actual simulations of the turbulent case, shows the ability of the hybrid DI/ANN model based on mass conservation error to improve the quality of the predictions and thus the robustness of ANNs. The increase in computational cost due to the hybrid model is acceptable as the DI is only applied in very localized regions in space and time.