Self-supervised generative models for crystal structures
Fangze Liu,
Zhantao Chen,
Tianyi Liu,
Ruyi Song,
Yu Lin,
Joshua J. Turner,
Chunjing Jia
Affiliations
Fangze Liu
Department of Physics, Stanford University, Stanford, CA 94305, USA; Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Corresponding author
Zhantao Chen
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Linac Coherent Light Source, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Corresponding author
Tianyi Liu
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Department of Chemistry, Stanford University, Stanford, CA 94305, USA
Ruyi Song
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA
Yu Lin
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA
Joshua J. Turner
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Linac Coherent Light Source, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Corresponding author
Chunjing Jia
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025, USA; Department of Physics, University of Florida, Gainesville, FL 32611, USA; Corresponding author
Summary: Inspired by advancements in natural language processing, we utilize self-supervised learning and an equivariant graph neural network to develop a unified platform for training generative models capable of generating inorganic crystal structures, as well as efficiently adapting to downstream tasks in material property prediction. To mitigate the challenge of evaluating the reliability of generated structures during training, we employ a generative adversarial network (GAN) with its discriminator being a cost-effective reliability evaluator, significantly enhancing model performance. We demonstrate the utility of our model in optimizing crystal structures under predefined conditions. Without external properties acquired experimentally or numerically, our model further displays its capability to help understand inorganic crystal formation by grouping chemically similar elements. This paper extends an invitation to further explore the scientific understanding of material structures through generative models, offering a fresh perspective on the scope and efficacy of machine learning in material science.
