Artificial neural networks implementation to predict the solution of nonlinear ODE system for the application to turbulent combustion modeling

Abstract

Computational fluid dynamics (CFD) simulations need a wide variety of computational resources. In order to obtain an accurate solution, one can use direct numerical simulation (DNS) as a CFD model but DNS is a computationally expensive approach, even with small problems. When it is required to model chemical reactions that occur in turbulent flow, the modeling becomes more challenging due to the complex interaction between turbulent flow and chemical kinetics. One of the widely used approaches to model this interaction is the eddy dissipation concept (EDC). Its use requires solving a system of stiff-nonlinear ordinary differential equations, which is computationally expensive as well. According to the literature, solving EDC model equations via artificial neural networks (ANNs) reduced significantly the computational time needed, but those ANNs were limited to a fixed time-step. The present research work investigates a novel approach to solve EDC model equations using ANNs with an adaptable time step instead of using the numerical traditional solver. Therefore, ANN with an adaptable time step has been implemented to solve the EDC model equations. Different data normalization techniques have been tested. The error of propagation for different time steps has been reported.