MBAvol: An Interpretable Descriptor Scheme for Crystal-Density Modeling of HEDM

Abstract

Machine learning (ML) offers an efficient route to predict properties of high-energy-density materials (HEDMs), but conventional models often have limited interpretability and out-of-sample accuracy. Here, we introduce the multibody approximation volume descriptor (MBAvol) for crystal-density (ρcryst) prediction. MBAvol represents molecular spatial information through physically motivated atomic and multibody volume terms derived from density functional theory and analytical fitting, thereby reducing the data required for accurate modeling. The resulting regression model achieves a mean absolute error of 0.026 g/cm3 and performs well for energetic cocrystals. SHAP analysis of the MBAvol terms provides quantitative, chemically interpretable structure–density relationships, from which three design guidelines were derived. Guided by these relationships, eight candidate HEDMs with high predicted ρcryst were identified. MBAvol therefore combines data efficiency, extrapolative capability, and interpretable feature representation, providing a practical framework for crystal-density prediction and structure-guided HEDM discovery.