Machine Learning Prediction of Crystal Structure Stability toward the Design of High-Entropy Oxides

Energy above the convex hull (E_hull) is a key thermodynamic criterion for assessing phase stability. However, the enormous computational cost required for the phase diagram construction hinders the prediction of E_hull, underscoring the need for data-driven approaches. Here, a hybrid framework integrating autoencoder with random forest classifier was proposed to effectively categorize crystal structures into stable, metastable, and unstable regimes according to E_hull thresholds, achieving an overall accuracy above 84%. More importantly, physically interpretable latent features associated with density, symmetry, and lattice were identified for stability prediction. Application to high-entropy oxides (HEOs) further demonstrates the effectiveness of the framework, revealing that structures with high configurational entropies and low cation radius mismatch are overwhelmingly classified as stable or metastable. Beyond confirming the dominant role of density and lattice feature in stability prediction, SHAP analysis further suggests that lager disparities in atomic thermal conductivities and the regulation of the magnetic moment by limited magnetic atoms play a critical role in governing the stability of HEOs structures. The interpretable and effective AE-RF algorithm developed in this work holds great potential for accelerating the discovery of novel HEOs and multicomponent materials.