Discovery of Quaternary Perovskites via Stacked Machine Learning and Generative Models
Abstract
This study explores machine learning-driven prediction of perovskite cathode materials for solid oxide fuel cells (SOFCs) by analyzing formation energy and bandgap properties. Addressing experimental challenges in material discovery, we developed a predictive framework using 602,000 compounds with feature-engineered atomic properties. Four algorithms (random forest, support vector machine, gradient boosting, and decision tree) and a generative adversarial network/variational autoencoders (GAN/VAE)-enhanced stacking model were evaluated through 100 iterations of 70-30 train-test splits. The stacking model achieved superior performance with 96% (±0.09%) tenfold cross-validation accuracy, significantly outperforming previous 60% benchmarks. Feature importance analysis identified key atomic properties guiding material selection. Through generative adversarial networks, we improved bandgap and formation energy predictions, ultimately screening 1,981 promising quaternary perovskites. This data-driven approach demonstrates substantial potential for accelerating SOFC cathode development by reducing experimental trial cycles through reliable computational pre-screening. The methodology establishes a robust foundation for machine learning applications in high-temperature material discovery, particularly in optimizing complex multi-property requirements for energy conversion technologies.