Prediction and Interpretability Analysis for Properties of Energetic Ionic Salts Based on Machine Learning

Abstract

As a significant subset of energetic materials, energetic ionic salts (EISs) have garnered considerable attention due to their potential application in the fields of explosives and propellants. To accelerate their development, this study establishes machine learning models for predicting key properties of EISs including density (ρ), enthalpy of formation (ΔHf), detonation velocity (DV), detonation pressure (P), and thermal decomposition temperature (Td). Using a customized descriptor set (CDS) and a dataset of 1,202 EISs, five ML algorithms were evaluated. The results show that optimal prediction model varies depending on target property: MLP performs best for ρ (R² = 0.91, MAE = 0.036 g·cm-3), SVR for P (R2 = 0.73, MAE = 1.86 GPa), KRR for ΔHf (R² = 0.85, MAE = 103.24 kJ·mol-1) and DV (R² = 0.69, MAE = 230.97 m·s-1), and GBR for Td. Prediction accuracy for ρ and ΔHf exceeded that of detonation properties and Td. Submodels based on nitro group attachment (C-NO₂ vs. N-NO₂) improved Td prediction. SHAP analysis revealed that oxygen balance, hydrogen, and nitrogen counts are universally important, alongside property-specific descriptors. This work demonstrates the potential of ML in predicting the key properties of EISs , thereby accelerating their structural design.