A New Artificial Neural Network Criterion for Identifying the Inherent Safety Process Conditions of Typical Autocatalytic Reactions

Abstract

In contrast to n-order reactions, autocatalytic reactions exhibit a characteristic where the products of the reaction further catalyze the process, which may significantly increase the risk of thermal runaway under certain conditions. In this work, a dimensionless mathematical model is established to characterize heterogeneous liquid–liquid autocatalytic reactions, with general applicability to any reaction order. Then, an artificial neural network (ANN) criterion is introduced to assess the thermal behavior of autocatalytic reactions, using seven representative features for characterization. To prevent overfitting, Bayesian regularization is applied during training, and particle swarm optimization (PSO) is used to determine the initial weights and thresholds. The results indicate that the established ANN criterion achieves high accuracies on training, validation, and test sets, all of which are more than 99%. The effectiveness of the ANN criterion is validated using a real autocatalytic reaction case. In practice, thermal behavior can be rapidly identified by evaluating reaction features at the current jacket temperature, offering an innovative strategy for identifying safe and efficient operating conditions for autocatalytic reactions.