Machine learning prediction of thermal sensitivity for energetic materials based on multi-source feature fusion

Thermal sensitivity, typically characterized by the thermal decomposition temperature (T_d), is a key parameter that reflects the thermal stability and safety of energetic materials, providing essential guidance for their engineering applications, storage, and transportation. However, traditional experimental determination methods suffer from long cycles, high costs, and significant hazards, rendering them unsuitable for efficient, large‑scale material screening. To address this bottleneck, this study systematically compares the predictive capabilities of topological features, calculated features, and their multi‑source fused features for estimating T_d. The results demonstrate that using only topological features in a Random Forest (RF) model yields an R² of 0.721, and the RMSE and MAE were 39.844 °C and 29.502 °C, respectively, whereas relying solely on calculated features gives considerably lower performance (R² = 0.462, RMSE = 42.391 °C, MAE = 34.836 °C). The fused features significantly improves prediction accuracy, achieving a test set R² of 0.799 and reducing RMSE and MAE to 31.585 °C and 24.443 °C, respectively. This enhancement is consistently observed across various machine learning models, including SVR, KRR, and AdaBoost, confirming the universal complementarity between the two feature types. Among the four models, RF combined with the SPXY sample partitioning method delivers the best overall predictive performance, providing a reliable computational tool and methodological framework for rapid and accurate T_d prediction of energetic materials.

All machine learning modeling and data processing were carried out in Python. Molecular topological features were generated using RDKit 2024.9.6. The crystal structures were obtained from the Cambridge Crystallographic Data Centre (CCDC), and subsequent first‑principles calculations were performed with the Cambridge Serial Total Energy Package (CASTEP) module in Materials Studio, employing the Perdew-Burke-Ernzerhof (PBE) functional within the generalized gradient approximation (GGA) with DFT‑D dispersion correction and norm‑conserving pseudopotentials. A plane‑wave cutoff energy of 830 eV was adopted, and geometry convergence criteria were set as follows: energy change < 5 × 10^-6 eV/atom, maximum force < 0.01 eV/Å, maximum stress < 0.02 GPa, and maximum displacement < 5 × 10^-4 Å. Single‑molecule calculations employed DMol³ with the same PBE functional and DNP basis set, with Grimme's D2 dispersion correction. The SCF convergence thresholds were set to 1 × 10^-5 Ha for energy, 0.002 Ha/Å for force, and 0.005 Å for displacement.