The reacting flow involves nonlinear partial differential equations which are mostly solved numerically for autocatalytic reaction between multi reactants. One-dimensional convection-diffusion-reaction equations are widely used for modeling reacting flows. This paper developed a method to solve the convection-diffusion-reaction equations using a neural network model, called the physics-informed neural network. It was applied to solve the nonlinear reacting flow for forward as well as inverse problems. To enhance accuracy and to eliminate non-physical oscillations, a derivative constraint of the spatial dimension was employed. In order to avoid the slow training speed caused by the derivative constraint, a new neural network connection structure is proposed. For validation, three different forward reacting flow problems involving steep fronts were analyzed, including the convection reaction, and steady and unsteady convection-diffusion-reaction systems. For the inverse problem, except for the kinetic parameters, the unknown convection velocity and diffusion coefficients were correctly predicted, demonstrating the capability of the proposed method for identifying reacting flow parameters. Moreover, under the guidance of the improved algorithm, the sensitivity of the reacting flow to certain kinetic parameter was investigated.
Read full abstract