3D Generative Modeling for Electrochemical Microstructures: A Comparative Study of Physics-Informed GANs and Discrete Latent Diffusion

Abstract

The performance of lithium-ion battery electrodes at a macro level is controlled by the electrode’s three-dimensional (3D) microstructure. However, as opposed to expensive 3D imaging, the macro-scale performance can be generated in a more scalable way through the creation of digital representations or “digital twins.” Current deep generative models are limited due to large 𝒪(𝑁3) memory requirements and constraints in enforcing physically correct information. This study provides a comparative analysis of three different generative approaches for creating multi-phase cathode NMC532 structures: a 2D-to-3D slice-based generative adversarial network (SliceGAN); a new physics-informed conditional generative adversarial network (pc-cGAN); and a discrete latent diffusion model (LDM). Results show that the SliceGAN is capable of achieving high levels of topological accuracy; however, the model does not enforce any constraints on how the material can be created, which limits its usefulness for inverse design. In contrast, the pccGAN effectively constrains the macroscopic properties of the outputted materials (such as porosity, tortuosity) using a 3D physics surrogate; however, its 𝒪(𝑁3) computational complexity restricts the number of voxels produced to 1283 and results in significant adversarial instabilities. The Discrete LDM uses a customized Dice-loss variational autoencoder (VAE) to map categorical representations of microstructures to a continuous latent space. By doing so, the model eliminates memory bottlenecks, enabling the production of highly stable volumes up to 2563. Although the conversion from continuous-to-discrete representations of the input produces minor amounts of smoothening near boundaries, the Latent Diffusion approach has the potential to become a highly scalable and robust framework for the application of artificial intelligence in materials design