CryinGAN: Design and evaluation of point-cloud-based generative adversarial networks using disordered interface structures

Generative models have received significant attention in recent years for materials discovery, but current efforts have mostly focused on generating simpler unit cells, and less attention has been given to advancing these models for more complex disordered materials. These models have also been developed by evaluating on the new, unverified materials being generated, resulting in limited metrics to evaluate model performance. In this work, we demonstrate how developing and evaluating generative models using a fixed set of disordered materials can improve these models for more complex materials, through direct comparisons between training and generated structures. Using a disordered Li₃ScCl₆-LiCoO₂ battery interface as our material system, we tested different point-cloud-based generative adversarial network architectures that further include bond distance information in the discriminator, instead of just atom coordinates. By analyzing the energy and structural features of the generated data, we were able identify reasons for the success/failure of design choices across architectures. We will show that our best performing architecture, Crystal Interface Generative Adversarial Network (CryinGAN), is capable of generating low-interface-energy structures with > 250 atoms that are energetically and structurally similar to the training structures.