Physics Guided Deep Learning for Generative Design of Crystal Materials with Symmetry Constraints

Discovering new materials is a long-standing challenging task that is crucial to the progress of human society. Conventional approaches based on trial-and-error experiments and computational simulations are labor-intensive or costly with success heavily depending on experts' heuristic knowledge. Recently, deep generative models have been proposed for generative design of materials by learning implicit knowledge from known materials datasets. However, these models are either applicable to a specific material system or the performance is low due to their failure to incorporate physical rules into their model training process. Here, we propose a deep learning based Physics Guided Crystal Generative Model (PGCGM) for efficient generative materials design with high structural diversity (up to 20 different space groups). The high performance of our model manifests its capability to capture and exploit the symmetric constraints of crystals and the pairwise atomic distance constraints among neighbor atoms. Using data augmentation and spatial atom clustering and merging, our PGCGM model increases the overall generation validity performance by more than 700\% compared to FTCP, one of the state-of-the-art structure generators and by more than 45\% compared to our previous CubicGAN model. The newly generated crystal materials also show higher quality in terms of atomic spatial distribution and composition diversity. We further validated the new crystal structures by Density Functional Theory (DFT) calculations. 1,869 materials out of 2,000 were successfully optimized, of which 39.6\% have negative formation energy and 5.3\% have energy-above-hull less than 0.25 eV/atom, indicating their thermodynamic stability and potential synthesizability. The 1,869 crystal structures have been deposited to the Carolina Materials Database \url{www.carolinamatdb.org}.