Machine Learning-Aided High-Throughput Virtual Screening of Novel Fused Nitrogen-Rich Heterocyclic Energetic Compounds

Energetic materials are crucial for defense and civilian applications, and fused nitrogen‑rich heterocyclic scaffolds are promising next‑generation energetic candidates due to their high energy density. However, traditional discovery methods suffer from the trade‑off between screening efficiency and accuracy, hindering high‑throughput exploration. In this study, we developed an integrated ML‑DFT pipeline for high‑throughput virtual screening of such compounds. A curated dataset of 4007 unique energetic compounds from EM Database v1.0 was used, covering six key properties: Density ( ρ ), Detonation Velocity ( D ), Detonation Pressure ( P ), Heat of Detonation ( Q ), Detonation Volume ( V ), Solid Phase Enthalpies of Formation ( ΔfHₘ° ), and. Six machine learning regression algorithms, Random Forest (RF), XGBoost, Support Vector Regression (SVR), Decision Tree (DT), K-Nearest Neighbors (KNN), and Ridge Regression were systematically optimized, and a task-specific ensemble model integrating Random Forest, XGBoost, and Decision Tree was constructed with R²>0.8 for key detonation properties. This model screened 10,000 derivatives from 5 fused heterocyclic scaffolds and 23 energetic substituents, identifying 20 top candidates. DFT validation at the B3LYP/6‑311 + G(d,p) level showed excellent consistency with ML predictions (most relative errors < 5%). The top candidate M10 exhibited exceptional performance: ρ  = 2.451 g/cm³, D  = 9.82 km/s, P  = 51.5 GPa. This work provides validated molecular targets and a transferable ML‑DFT framework, balancing efficiency and accuracy to accelerate the discovery of next-generation energetic materials.