Perovskite Formation Energy and Bandgap Prediction: A Graph Neural Network and Transfer Learning Framework

Key Points

• The aim is to predict the formation energy and bandgap of perovskite materials using advanced machine learning techniques.
• Developed a framework combining graph neural networks with transfer learning.
• Utilized data from the Computational Materials Repository for training models.
• Employed MEGNet, SchNet, and CGCNN for predicting formation energy.
• MEGNet demonstrated the highest prediction accuracy for formation energy with a MAE of 0.070 eV/atom.
• MEGNet improved the determination coefficient R² to 0.986.
• Transfer learning reduced the MAE for bandgap predictions from 0.42 eV to 0.34 eV.