An improved approach towards more robust deep learning models for chemical kinetics

Abstract

The application of artificial neural network (ANN) models in replacement of numerical integration solvers for ordinary differential equations (ODE) has attracted increasing interest in recent years owing to its high computational speed potential and approximation capabilities. However, regarding its application for the calculation of chemical kinetics in combustion simulations, there are challenges over its suitability in wide-range reaction conditions, error propagation over excessive time steps, and its performance in cases with high nonlinearity (especially at elevated pressures). In this study, a deep learning model was successfully developed as a fast alternative to the numerical integration of the detailed mechanism of hydrogen for time-resolved ignition simulations at pressures up to 45 atm, which is featured by the abruptness of ignition and long ignition delays. More importantly, some methodological improvements are proposed and discussed in an effort to build more robust models and to extend the applicable range of the deep learning model without sacrificing model compactness. These techniques include the employment of continuously differentiable activation functions, rate-distinguishing classification, and perturbed simulation for data generation. This study shows that these techniques can unlock further potentials of ANN in making reliable predictions in practical applications. The fair comparison between these methods and conventional ones in this study also enables a deeper understanding of the effect of data distribution and network architecture on the performance of ANN models in solving stiff ODEs.