Advanced Machine Learning Models for Predicting Perovskite Formation Energy: Ensemble Methods and Neural Networks

The prediction of formation energy in perovskite materials (ABX₃) is crucial for accelerating material discovery in energy-related applications. This study applied six machine learning models—Extra Trees, Gradient Boosting Regressor (GBR), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), Deep Neural Networks (DNN), and a hybrid Gradient Boosting Recurrent Neural Network (GBRNN)—to model the complex relationships between formation energy and elemental properties such as electronegativity, atomic radius, and ionization energy. The models were trained and evaluated on computational datasets using metrics including R², Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). Robustness was ensured through 5-fold and 10-fold cross-validation, while Bayesian optimization fine-tuned hyperparameters such as learning rate, tree depth, batch size, and dropout rate. SHAP analysis identified bond variability and B-site cation properties as the most significant predictors, providing insights into structural characteristics influencing material stability. Among the models, the GBRNN hybrid demonstrated the best performance, achieving an R² of 0.9818 and effectively capturing complex non-linear interactions in the data. This study underscores the value of combining machine learning with feature interpretability and hyperparameter tuning to predict formation energy with high accuracy, offering actionable insights for designing stable perovskite materials and optimizing their properties for energy applications.