Center-Environment Deep Transfer Machine Learning for Crystal Structures: Spinel to Perovskite Oxides.

Scientists have developed a deep transfer learning method to predict the stability of perovskite oxides. Using a large computational dataset of spinel oxides, the team trained a deep neural network (DNN) source domain model with Center-Environment (CE) features. They then fine-tuned this system by learning a small dataset of 855 perovskite oxide structures. The resultant transfer learning model successfully predicted the formation energy of 1314 thermodynamically stable perovskite oxides. Of these, 144 oxides have already been synthesized experimentally, with 10 predicted computationally by other literature, while 859 have only been discovered through this study. This approach promises a more cost-effective route to screening the expensive high-throughput computational screening required for data-driven materials design. The findings suggest that transfer machine learning is a valuable tool for predicting stable novel perovskite oxides with a view to exploring potential renewable energy and electronic materials applications.