Intelligent high-energy molecules design for solid propellants via interpretable machine learning and latent space optimization

Intelligently designing molecules based on composite material properties is the ideal goal across many areas of materials science. For solid propellants, where high-energy molecules critically affect performance, property-driven molecular design is highly desirable but remains unrealized. To this end, we present an AI workflow that integrates property prediction, feature attribution, and generative modeling to enable interpretable molecular design for high-energy solid propellants. A curated dataset of over 16,000 molecules supports both accurate multi-target prediction models and a generative framework. The predictor attains high accuracy for Hydroxyl-terminated polybutadiene (HTPB) system formulations (R² ≥ 0.996) and generalizes to Glycidyl Azide Polymer (GAP) and NitrateEster Plasticized PolyetherPropellant (NEPE) systems (R² ≥ 0.970). SHapley Additive exPlanations (SHAP) analysis highlights oxygen balance (OB_CO > −10 %) and nitrogen content (35–45 wt%) as key factors. Guided by these insights, latent space optimization using a Junction Tree Variational Autoencoder and Bayesian optimization yields over 200 novel high-energy molecules with predicted I_sp above 265 s. This work demonstrates an interpretable pipeline for accelerating high-energy molecules discovery and propellant formulation design.