Rational Design of Single-Phase High-Entropy Oxides via Large Language Model Data Mining and Explainable Machine Learning

-score of 86% in a seven-class classification problem, and the best-performing database variant combined primary and statistical features. This optimal database representation was then used to train a neural-network binary classifier to distinguish perovskite from nonperovskite compositions, achieving 97.9% classification accuracy on the test set, whereas the Goldschmidt tolerance factor reached only 67.3% on the same data. These results indicate that the proposed methodology can support the design of HEO compositions with target properties and substantially outperforms traditional descriptor-based approaches. Furthermore, SHAP (SHapley Additive exPlanations) analysis revealed that high-entropy perovskite phase stability is governed by a critical interplay between geometric factors, such as the sum of cation radii, and electronic descriptors, including Sanderson electronegativity and atomization enthalpy. Overall, these findings demonstrate that LLM-driven data mining can overcome data bottlenecks and enable the discovery of physical design rules for complex, multicomponent ceramics.