Crystal Structure Prediction Using Generative Adversarial\nNetwork with Data-Driven Latent Space Fusion Strategy

Crystal\nstructure prediction (CSP) is an important field of material\ndesign. Herein, we propose a novel generative adversarial network\nmodel, guided by a data-driven approach and incorporating the real\nphysical structure of crystals, to address the complexity of high-dimensional\ndata and improve prediction accuracy in materials science. The model,\ntermed GAN-DDLSF, introduces a novel sampling method called data-driven\nlatent space fusion (DDLSF), which aims to optimize the latent space\nof generative adversarial networks (GANs) by combining the statistical\nproperties of real data with a standard Gaussian distribution, effectively\nmitigating the “mode collapse” problem prevalent in\nGANs. Our approach introduces a more refined generation mechanism\nspecifically for binary crystal structures such as gallium nitride\n(GaN). By optimizing for the specific crystallographic features of\nGaN while maintaining structural rationality, we achieve higher precision\nand efficiency in predicting and designing structures for this particular\nmaterial system. The model generates 9321 GaN binary crystal structures,\nwith 16.59% reaching a stable state and 24.21% found to be metastable.\nThese results can significantly enhance the accuracy of crystal structure\npredictions and provide valuable insights into the potential of the\nGAN-DDLSF approach for the discovery and design of binary, ternary,\nand multinary materials, offering new perspectives and methods for\nmaterials science research and applications.