A Machine Learning Tool for Thermal Runaway Prediction of Chemical Reactors

The thermal runaway of chemical reactors has been a pivotal issue for safe operation in the chemical engineering and related fields. In the past, mathematical criteria were employed to predict the onset of thermal runaway as well as the boundary between safe and risky operations. Owning to the time-consuming computation and mathematical threshold, however, these criteria may fail in rapidly predicting thermal runaway for extensive cases. In the present work, a machine learning (ML) tool was applied to analyze 30,000 cases of reactor thermal behavior. A new ML-based model including a criterion (η) was proposed to describe the thermal behavior of chemical reactors. It was found that the Random Forest algorithm implemented in this model provides a reliable prediction for both the onset (η=1) and intensity of reactor thermal runaway. Using the proposed criterion, the non-runaway scenario (η<1) is classified into three areas as follows, highly risky (0.9<η<1), intermediately risky (0.5<η<0.9) and relatively safe (0<η<0.5). The present ML-based model with a low mathematical threshold is a promising start toward the rapid evaluation of chemical reactor thermal behavior in practice.