Abstract
Accelerating the discovery of novel crystal materials by machine learning is crucial for advancing various technologies from clean energy to information processing. The machine-learning models for prediction of materials properties require embedding atomic information, while traditional methods have limited effectiveness in enhancing prediction accuracy. Here, we proposed an atomic embedding strategy called universal atomic embeddings (UAEs) for their broad applicability as atomic fingerprints, and generated the UAE tensors based on the proposed CrystalTransformer model. By performing experiments on widely-used materials database, our CrystalTransformer-based UAEs (ct-UAEs) are shown to accurately capture complex atomic features, leading to a 14% improvement in prediction accuracy on CGCNN and 18% on ALIGNN when using formation energies as the target, based on the Materials Project database. We also demonstrated the good transferability of ct-UAEs across various databases. Based on the clustering analysis for multi-task ct-UAEs, the elements in the periodic table can be categorized with reasonable connections between atomic features and targeted crystal properties. After applying ct-UAEs to predict formation energy in hybrid perovskites database, we realized an improvement in accuracy, with a 34% boost in MEGNET and 16% in CGCNN, showcasing their potential as atomic fingerprints to address the data scarcity challenges.