Machine learning-accelerated efficient screening and design of energetic material molecules

Energetic materials play a crucial role in military and aerospace applications. However, the discovery and synthesis of novel energetic compounds still largely rely on traditional trial-and-error approaches, which severely hinder the development of novel energetic materials. This study focuses on the prediction of a key thermodynamic property of energetic materials—heat of formation (HOF) and proposes a machine learning structure-property relationship model that integrates active learning strategies with SMILES-based molecular feature representation. A dataset containing 1447 gas-phase energetic molecules was constructed based on the high-accuracy G4 quantum chemical method, and 93 effective SMILES descriptors were extracted to establish a preliminary model for gas-phase HOF using linear model. Subsequently, the model was applied on the systematical prediction of 221738 potential energetic molecules retrieved from the PubChem database, enabling the screening of candidates with superior explosive performance. For samples with high prediction errors, an active learning strategy was implemented to iteratively refine the model parameters, significantly improving prediction accuracy. Validation on classical energetic molecules illustrated excellent predictive performance, highlighting the model’s strong generalization capability. Finally, 20 candidate molecules with a TNT equivalent power index exceeding 2.0 were screened, most of which are new to the existing reservoir of known energetic materials. These results underscore the potential of this proposed approach in accelerating the discovery of high-performance energetic materials and offer a new strategy for the development of next-generation energetic compounds.