Machine Learning-Driven High-Throughput Discovery of Energetic Materials Based on Regioisomeric Pyrazole Scaffolds

The energy-safety trade-off remains the central challenge in energetic materials design. This study constructs a virtual library of 25,417 derivatives based on two regioisomeric pyrazole scaffolds and trains predictive models using Density Functional Theory (DFT) data from 413 molecules. Unlike conventional approaches, we systematically benchmarked three descriptor families and identified RDKit global physicochemical descriptors as superior to Morgan fingerprints, revealing that global molecular properties govern detonation performance in this regioisomeric system. Among six evaluated machine learning algorithms, Extreme Gradient Boosting (XGBoost) was identified as the optimal base model. Building on this, we developed a multi-task XGBoost framework that jointly trains four physically coupled properties, further improving average R2 from 0.891 to 0.927. Critically, SHAP analysis reveals that the amino group (fr_NH2) exerts a negative net contribution to detonation velocity—direct evidence of the energy-safety trade-off—and enables regioisomer-specific design rules. Screening the remaining 25,004 molecules and validating the top 20 candidates by blind DFT calculations yield mean relative errors of 3.02% for D and 5.76% for P, with the best candidate below 2%. This interpretable, multi-task screening paradigm offers a generalizable platform for accelerated energetic materials discovery.