Forward and reverse problems of methane dehydro-aromatization based on physics-informed neural network
Research on solving the forward and reverse problems of chemical reaction kinetics modeling can help to gain a deeper understanding of reaction mechanisms and reduce experimental costs.This study took the one-dimensional packed bed methane dehydro-aromatization(MDA)as an example and used a physics-informed neural network to couple the chemical reaction mechanism equations into the loss function.In this way,a solution framework for reaction kinetics modeling and parameter inversion was constructed.Firstly,the optimal neural network hyperparameters were determined by solving the forward problem.The results showed that the constructed model had good predictive performance in solving the MDA reaction kinetics model,with training error and extrapolation error L2 of 0.19%and 0.95%,respectively.Based on this,the rate constants of MDA were inverted using labeled data under 0,0.1%,and 0.3%Gaussian noise,and the predicted values obtained from training had a relative error within 0.5%of the true values,demonstrating the ability of physics-informed learning to perform inversion for unknown kinetic parameters under low-quality data.
methane dehydro-aromatizationphysics-informed neural networkreaction kinetic modelinverse problem
国家科技期刊平台
NSTL
万方数据
