Machine Learning Accelerated Analysis of Chemical Reaction Networks for Gas-Phase Reaction Systems

Abstract

Chemical reaction networks (CRNs) serve to describe the behavior of complex chemical reaction systems. Analyzing CRNs of a reactive system requires kinetic data that are typically obtained by time-consuming experiments or computational chemistry. Machine learning (ML) has emerged as a promising approach for rapid property prediction based on historical data. However, the accuracy of ML model predictions in kinetics remains a limitation for their application in CRN analysis. In this study, we integrate the cross-attention mechanisms in neural networks and CRN sensitivity and uncertainty analysis to enable the practical application of the ML models in reliable gas-phase CRN analysis. Specifically, a message-passing neural network (MPNN) architecture along with a cross-attention mechanism (CA-MPNN) was developed for accurate prediction of the reaction rate constants with prediction uncertainty. CA-MPNN model outperformed the conventional deep neural network architectures on most of the reaction property data sets. We combined reaction network sensitivity analysis and ML prediction uncertainty analysis to identify influential reactions with high-level uncertainty of the predicted rate constant, which are further calibrated using high-accuracy quantum chemistry methods to mitigate the problem of inaccurate machine learning predictions. Compared with the traditional workflow, this framework significantly reduces up to 80% computational cost to construct a reliable CRN in the demonstrated gas-phase pyrolysis and combustion applications.