Descriptors applicability in machine learning-assisted prediction of thermal decomposition temperatures for energetic materials: Insights from model evaluation and outlier analysis

Abstract

Machine learning is an emerging approach to predict thermal decomposition temperature in the field of energetic materials, while an assessment of the descriptor applicability is still lacking. In this work, we have systematically established 5 general descriptor sets for 1091 compounds and combined them with 9 algorithms to construct a suite of predictive models with mean absolute error ranging 41–29 K, which is comparable to the cutting-edge endeavors. Our study emphasizes the significant influence of multi-level structural interactions on the thermal stability and decomposition of energetic materials, contributing insights conducive to the development of corresponding descriptors.