Deep multi-field network for physical and concentration field prediction of TDM reactor under different carbon particle distributions

Abstract

In thermal decomposition of methane (TDM) reactors, the injection of carbon particles and their distribution have significant influence on field distributions, thereby impacting the yield and wall deposition rate. It is computationally intensive to analyze such influences numerically by repetitively solving the mechanism model. Therefore, we developed a multi-field network (DMN) to predict the physical and concentration fields of a TDM reactor under different inlet distributions of injected carbon particles. A set of operator networks was trained using small amounts of data generated from the mechanism model, and then, it was integrated with a multi-hidden layer neural network (NN) to form an improved physics-informed NN, which is regarded as a DMN. By comparison, the proposed DMN can predict multiple fields under any different particle distribution with higher accuracy than the traditional NN and higher efficiency than solving the mechanism model. The proposed DMN can also be used for field reconstruction using observational data.