Multi-Body Approximation Featurization to Broaden Machine Learning Application Range for Crystal Density Modelling

Machine Learning (ML) is pivotal in forecasting the physicochemical properties of high-energy-density materials (HEDMs). However, the limited out-of-sample predictive capacity and interpretability of conventional models have constrained advancements beyond the performance boundaries defined by the training datasets. This research introduces a novel molecular descriptor, Multi-body Approximation (MBA), which enhances the prediction of crystal density (ρcryst)-a critical parameter for detonation performance-by incorporating molecular spatial information into atomic groups. Analytical functions for MBA volume terms were developed through Density Functional Theory simulations and theoretical derivation, substantially reducing the data requirements for constructing accurate ML models. The resulting general regression model achieved exceptional accuracy, with a mean absolute error (MAE) of 0.026 g/cm³. Furthermore, the MBA volume descriptor demonstrated excellent predictive capability for more complex co-crystal structures. Leveraging the SHAP explainer, we derived three interpretable design criteria for high-ρcryst materials and subsequently identified eight high-ρcryst candidates. These results demonstrate that the MBA volume descriptor effectively integrates chemical and physical principles into a mathematical framework, achieving multiple advanced functionalities: compatibility with small datasets, out-of-sample extrapolative capability, and intuitive model interpretability. We anticipate that extending MBA volume descriptors to other HEDM metrics, coupled with probabilistic structure generation models, holds significant potential to revolutionize HEDM design.