Data-driven 2D grain growth microstructure prediction using deep learning and spectral graph theory

Abstract

In this paper, we present an alternative method to grain growth simulations. Traditional grain growth algorithms can be computationally expensive, especially when considering anisotropic grain boundary (GB) properties. The new Semi-Stochastic Grain Growth Prediction (SSGGP) model consists of two main components: a statistical evolution model that predicts the evolution of the GB network spectrum and a conditional diffusion model that generates grain growth morphologies at different time steps. These models are trained on a dataset Niño and Johnson (2024) that contains thousands of microstructures obtained from anisotropic grain growth simulations. We test the effectiveness of our model by comparing microstructure statistics (e.g., grain size distribution, orientation distribution function (ODF), misorientation distribution function (MDF), and GB energy distribution) with those obtained from grain growth simulations. The results indicate that the SSGGP model shows good agreement in terms of these statistics. Moreover, once trained, the SSGGP is almost ten times faster in obtaining the evolved state of a microstructure. We also find evidence for self-similarity of the GB network during steady-state normal anisotropic grain growth.