Comparative Analysis of Classical Machine Learning and Graph Neural Network Models for Perovskite Property Prediction

Jirui Jin, Somayeh Faraji, Bin Liu, and Mingjie Liu

University of Florida

Perovskite materials, renowned for their versatility and outstanding properties, pose challenges in discovering optimal candidates due to the vast compositional space, resource-intensive synthesis, and limited understanding of structure-property relationships. Data-driven machine learning (ML) methods offer promise in navigating this complexity and expediting material discovery; however, the trade-off between accuracy and efficiency across different ML models for predicting perovskite properties is not well understood. In this study, we conducted a comprehensive and consistent assessment of various ML models for predicting the formation energy (Eg) and bandgap (Eg) of perovskite materials. We designed a solid pipeline to extract perovskite structures from distinct datasets based on octahedral lattice motif. Benchmarking classical ML algorithms (CML) against graph neural network (GNN) models across three datasets revealed the GATGNN model as the top performer, balancing exceptional prediction accuracy and computational efficiency. Furthermore, we investigated the impact of data size on model performance, emphasizing the need for over 1000 data points for optimal prediction accuracy. Additionally, through SHAP value analysis, we determined the group number of atom X, the covalent radius of atom A, and the electronegativity of atom B as the top three most important features for Ef prediction, whereas the electron affinity of atom X and the valence electrons of atom B emerged as the top two for Eg prediction. Our study establishes a standardized benchmark for evaluating various ML models across diverse datasets of perovskite materials, facilitating future applications in materials science, particularly in model selection and the advancement of perovskite materials.