Interpretable machine learning for predicting electric spark sensitivity of energetic compounds via molecular and instrument descriptors

Accidental ignition of energetic compounds by electrostatic discharge poses significant safety risks, highlighting the need for precise prediction of electric spark sensitivity. Here we develop a novel interpretable machine learning (ML) model using a dataset of 211 measurements for 160 CHON-based energetic compounds (including ionic salts), which are tested with both RDAD and KTTV instruments. An instrument descriptor and 19 molecular descriptors derived from chemical structure representations served as input features to four algorithms: Random Forest (RF), Support Vector Regression, Back-Propagation Neural Network, and Multilayer Perceptron. Hyperparameters for each model were optimized via a genetic algorithm under five-fold cross-validation. The RF model exhibited the highest performance (R 2 = 0.923) with the lowest predictive error, thus selected for subsequent interpretation. SHapley Additive exPlanations (SHAP) analysis further identified key molecular descriptors, including minimum partial charge, oxygen balance, and octanol-water partition coefficient. Our ML model covers a broader range of energetic materials and generalizes across two common testing instruments. Additionally, it employs accessible descriptors that require no costly simulations or complex calculations. This data-driven methodology provides a rapid, accessible and cost-effective framework for predicting electric spark sensitivity of diverse energetic compounds, supporting the design and testing of safer energetic compounds, while offering insights for future research on other sensitivities. • Developed an interpretable Random Forest model for predicting electric spark sensitivity of energetic compounds, achieving the best performance with an R 2 of 0.923. • Collected 211 measurements covering 160 CHON-based energetic compounds (including ionic salts) tested with two commonly adopted RDAD and KTTV instruments, improving model robustness and practical relevance. • Built the model using only molecular and instrument descriptors, enabling a rapid, low-cost, and simulation-free alternative to traditional prediction methods.