Pioneering Heat of Combustion of Energetic Materials via Data‐Based Approaches

Energetic materials play a vital role in military, aerospace, and industrial applications, but their synthesis and evaluation are costly, hazardous, and time‑consuming. This study aims to develop accurate and interpretable machine learning (ML) models for predicting the heat of combustion of energetic materials based on molecular descriptors. A curated dataset of 325 compounds with experimentally measured heats of combustion was assembled from multiple literature sources, including descriptors such as molecular weight (MW) and the presence of key functional groups. A Gradient Boosted Decision Trees (GBDT) model was optimized using the following four hyperparameter search strategies: Gaussian process optimization (GPO), evolutionary strategies (ES), Bayesian probability improvement (BPI), and batch Bayesian optimization (BBO). The models were trained and validated through five‑fold cross‑validation followed by a 90/10 hold‑out split, and their performance was assessed using R 2 , mean squared error (MSE), and average absolute relative error (AARE%). Among the tested approaches, GBDT–GPO demonstrated the highest generalization capability, achieving the lowest test‑phase error (MSE = 70 571, AARE% = 5.456, R 2 = 0.947), followed by ES and BBO, while BPI showed slower convergence and larger prediction errors. To rigorously verify generalization and eliminate any risk of data leakage, the optimized models were further evaluated on a newly constructed, strictly independent experimental dataset of 23 energetic materials. The GBDT–GPO model maintained its superior performance on this entirely unseen hold‐out set ( R 2 = 0.988), unequivocally confirming its robust predictive capability. Sensitivity analysis identified azide groups, MW, and aromatic rings as the most influential descriptors, a finding supported by SHAP feature importance and contribution plots, which were consistent with established thermochemical reasoning. These results underline the value of combining advanced optimization with interpretable ML for guiding the design of high‑performance energetic compounds, while reducing experimental risks and costs through reliable in silico evaluation.