Interpretable machine-learning models for predicting elasticity and ductility of inorganic materials
Phys. Rev. Materials 10, 073602 – Published 7 July, 2026
DOI: https://doi.org/10.1103/3c4b-phhy
Abstract
Understanding and predicting mechanical properties such as elasticity and ductility are critical for the design of structural materials. Here, we develop a generalizable machine-learning framework to assess elastic properties and brittle-ductile tendencies across a chemically diverse space of inorganic compounds, including both binary and ternary systems. By integrating stacked ensemble learning and symbolic regression, we achieve strong predictive performance (e.g., for bulk modulus, for shear modulus, and ductile/brittle classification accuracy of 88%)—notably high for a general-purpose model spanning diverse chemistries. We also derive simple closed-form expressions using elemental and structural descriptors, enabling quick property estimation in practical settings. Finally, we apply the models to out-of-training-set compounds, including quaternary systems, and demonstrate their effectiveness in real applications. Relying only on several key basic descriptors, the physically interpretable models developed here are particularly valuable for quick screening workflows—especially for experimentalists without access to ab initio simulations.
