Abstract: Thermal stability is a critical property in the study of energetic materials, with the decomposition temperature serving as a central metric for evaluating their stability. In this study, a novel composite descriptor system was suggested, and a high-performance prediction model for the decomposition temperature of energetic materials was constructed. First, molecular structure descriptors based on group contribution were proposed. Then, bond dissociation energy (BDE) was introduced as a key supplementary parameter to quantify the effect of bond strength on decomposition temperature. Finally, RDKit descriptors were generated using the RDKit software, ultimately integrating them into a multidimensional feature set. This feature set was submitted to random forest (RF), support vector machine (SVM), and partial least squares (PLS) individually to construct multiple prediction models and conduct a systematic comparison. Among them, the best results were obtained using the model built by RF. Its prediction performance was superior to the results reported in the literature, indicating that the proposed composite descriptors can effectively capture the key factors affecting the decomposition temperature. To further interpret the model and identify critical influencing factors, SHAP visualization technology was employed to analyze the optimal model, thereby providing valuable data-driven insights into the thermal stability mechanisms of energetic materials.

Key words: Decomposition temperature, Molecular group descriptor; QSPR, Random Forest