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

Area of Science:

• Materials Science
• Computational Materials Science
• Data Science

Background:

• Rational design of high-entropy oxides (HEOs) is limited by a lack of structured property data in scientific literature.
• Developing predictive models for HEO single-phase stability is crucial for materials discovery.

Purpose of the Study:

• To create an end-to-end materials informatics framework for predicting single-phase stability in high-entropy oxides.
• To leverage large language models (LLMs) for data extraction and machine learning for property prediction.

Main Methods:

• Utilized LLM agents (gpt-oss-120b) to extract composition, phase, and synthesis data from unstructured abstracts with 96% accuracy.
• Trained multiclass classification models (XGBoost achieved 86% F1-score) to distinguish HEO crystal structures.
• Developed a neural network binary classifier, achieving 97.9% accuracy in predicting perovskite stability, outperforming traditional methods.

Main Results:

• An LLM-based agent successfully generated a structured database from scientific abstracts, inferring cation proportions.
• An XGBoost classifier distinguished seven HEO crystal structures with an 86% F1-score.
• A neural network classifier accurately predicted perovskite stability (97.9%), significantly outperforming the Goldschmidt tolerance factor (67.3%).

Conclusions:

• The proposed LLM-driven data mining and machine learning framework effectively overcomes data bottlenecks in HEO research.
• This methodology enables the design of HEO compositions with desired properties and discovers physical design rules.
• SHAP analysis revealed key geometric and electronic factors governing perovskite phase stability in HEOs.