Oral: An accelerated prediction of the mechanical properties of ABX₃ perovskites using interpretable machine learning models

Two techniques were assessed to develop an interpretable machine learning model that accelerates the prediction of the mechanical characteristics (bulk, shear, and Young’s moduli) of ABX₃ perovskites: (1) integrating elemental and DFT-derived features, and (2) utilising elemental features only. Three ensemble learning techniques CatBoost, Random Forest, and XGBoost were trained on a dataset of ABX₃ perovskite samples. Pearson Correlation Coefficient was used for the feature selection process. SHapley Additive exPlanations (SHAP) was adapted to provide physical insights into the model's interpretability. For the machine learning models including both elemental features and Density Functional Theory (DFT) derived features, it was observed that Random Forest achieved an R² of 0.965 in predicting the bulk modulus while XGBoost demonstrated superior performance in predicting shear and Young's moduli, with R² values of 0.967 and 0.974, respectively. SHAP analysis identified the elastic constants C₁₁ and C₄₄ as essential for predicting these moduli. Though the inclusion of these DFT-derived features enhances predictive accuracy, they can be computationally intensive and restrict scalability. In addressing this problem, an experiment on a machine learning model utilising solely elemental features was performed. For this experiment, Random Forest has an R² of 0.999 for moduli predictions, while CatBoost demonstrated competitive performance with an R² of 0.975 for bulk modulus and R² values of 0.993 for both shear and Young’s moduli. Based on the SHAP analysis, key factors include the covalent radii of elements B and X, melting temperature, and the electron affinity of element B, all of which greatly impact and accelerate the moduli predictions. This research highlights the potential of using simple, accessible elemental properties in expediting the discovery and enhancement of pressure-resistant perovskites across a broad chemical composition consistent with the objectives of materials informatics to facilitate material design and discovery.