The paper presents a method for solving the modeling of chemical kinetics that allows for quickly and with sufficient accuracy approximating combustion processes of a hydrogen-air mixture using neural networks.
The UNET-type neural network model was successfully applied to the problem of predicting deterministic multidimensional time series, namely, modeling chemical combustion processes described by a stiff system of ordinary differential equations.
We were able to train a compact model that can approximate changes in the concentrations of substances in a mixture during chemical reactions with a high degree of accuracy sufficient to obtain a prediction for hundreds or even thousands of integration steps ahead, while taking an order of magnitude less calculation time than numerical integration.
Preliminary experiments have shown that in most cases, the molar densities of substances change very slow in the beginning. In the present work, we propose to do a logarithmic renormalization of the data so that small changes near zero become more noticeable. In addition to logarithmic scaling, standard data scaling was applied. Also, we have included multi-step predictions in recurrent mode in the learning process itself. All these predictions are involved in the construction of the loss function during training.
