Machine learning-assisted quantitative prediction of thermal decomposition temperatures of energetic materials and their thermal stability analysis

In this study, machine learning (ML)-assisted regression modeling was conducted to predict the thermal decomposition temperatures and explore the factors that correlate with the thermal stability of energetic materials (EMs). The modeling was performed based on a dataset consisting of 885 various compounds using linear and nonlinear algorithms. The tree-based models established demonstrated acceptable predictive abilities, yielding a low mean absolute error (MAE) of 31 degrees C. By analyzing the dataset through hierarchical classification, this study insightfully identified the factors affecting EMs' thermal decomposition temperatures, with the overall accuracy improved through targeted modeling. The SHapley Additive exPlanations (SHAP) analysis indicated that descriptors such as BCUT2D, PEOE_VSA, MolLog_P, and TPSA played a significant role, demonstrating that the thermal decomposition process is influenced by multiple factors relating to the composition, electron distribution, chemical bond properties, and substituent type of molecules. Additionally, descriptors such as Carbon_contents and Oxygen_Balance proposed for characterizing EMs showed strong linear correlations with thermal decomposition temperatures. The trends of their SHAP values indicated that the most suitable ranges of Carbon_contents and Oxygen_Balance were 0.2--0.35 and -65---55, respectively. Overall, the study shows the potential of ML models for decomposition temperature prediction of EMs and provides insights into the characteristics of molecular descriptors.