CryinGAN: Design and evaluation of point-cloud-based generative adversarial networks using disordered materials $-$ application to Li$_3$ScCl$_6$-LiCoO$_2$ battery interfaces
Generative models have received significant attention in recent years for materials science applications, particularly in the area of inverse design for materials discovery. While current efforts have mainly focused on bulk materials with relatively small unit cells, the possibility of generative models for more complex, disordered materials would significantly extent the current capabilities of generative modeling. Generative models would also benefit from better ways to assess their performance. They are typically evaluated on the new, unverified materials being generated, which give limited metrics for evaluation. In this work, we design and evaluate models intended for disordered interface structure generation. Using a disordered Li$_3$ScCl$_6$-LiCoO$_2$ battery interface, we tested different point-cloud-based generative adversarial network architectures that further include bond distance information in the discriminator, rather than only atomic coordinates. By working with a fixed material system, we evaluated the model performance through direct comparisons between training and generated structures. The best performing architecture, Crystal Interface Generative Adversarial Network (CryinGAN), uses two separate 1D convolutional discriminators, one that accepts coordinates and another that accepts bond distances as input. We demonstrate that CryinGAN is able to successfully generate low-interface-energy structures for systems with > 250 atoms, in which the generated interfaces are structurally similar to the training structures. This study highlights the capabilities of a relatively simple generative model in generating large disordered materials, and discusses the limitations of the point cloud representation. Insights are provided to help guide the development of future generative models that are useful to not just disordered, but also ordered materials.
PDF Abstract
