Interpretable machine learning-assisted screening of perovskite oxides.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Perovskite oxides are crucial for energy storage and conversion technologies.
• Traditional screening methods (experimental and DFT) are time-consuming and costly.

Purpose of the Study:

• To develop interpretable machine learning models for predicting perovskite oxide stability.
• To accelerate the identification of stable perovskite oxides from vast virtual combinatorial libraries.

Main Methods:

• Constructed classification and regression models to predict thermodynamic stability and energy above the convex hull (Eh).
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.
• Utilized features such as highest occupied molecular orbital energy, elastic modulus, and ionic radius.

Main Results:

• A classification model achieved high accuracy (0.919), precision (0.937), F1-score (0.932), and recall (0.935).
• Successfully screened over 682,000 stable perovskite oxides from more than 1.1 million virtual combinations.
• A regression model predicted Eh values with a coefficient of determination of 0.916, showing good agreement with DFT calculations.

Conclusions:

• Interpretable machine learning offers an efficient alternative to conventional methods for perovskite oxide discovery.
• Key material descriptors like highest occupied molecular orbital energy and elastic modulus are vital for stability prediction.
• The developed models significantly expedite the search for novel perovskite materials for energy applications.