The use of machine learning methods for accelerating the design ofcrystalline materials usually requires manually constructed feature vectors orcomplex transformation of atom coordinates to input the crystal structure,which either constrains the model to certain crystal types or makes itdifficult to provide chemical insights. Here, we develop a crystal graphconvolutional neural networks (CGCNN) framework to directly learn materialproperties from the connection of atoms in the crystal, providing a universaland interpretable representation of the crystals structure. Our method achievesthe same accuracy as DFT for 8 different properties of crystals with variousstructure types and compositions after trained with $10^4$ data points.Further, our framework is interpretable because one can extract thecontributions from local chemical environments to global properties. Using anexample of perovskites, we show how this information can be utilized todiscover empirical rules for materials design.
展开▼
