Inductive determination of reaction–diffusion model parameters via dislocation pattern recognition using a convolutional neural network
Tatchaphon Leelaprachakul, Hiroyuki Shima, Takashi Sumigawa, Yoshitaka Umeno · Computational Materials Science · 2025
Dislocation patterning under cyclic loading is a hallmark of microstructural evolution in crystalline materials. The Walgraef–Aifantis (WA) model captures these phenomena through a set of nonlinear reaction–diffusion equations, yet the inductive determination of its parameters from observed patterns remains a significant challenge. This study presents a data-driven framework that leverages convolutional neural networks (CNN) to predict key WA model parameters, accounting for anisotropic diffusion, directly from simulated dislocation structures. A dataset of over 13,824 patterns was generated via numerical simulations under varied WA parameters. The CNN model demonstrates high accuracy in multi-parameter regression, enabling top-down inference of loading conditions from microstructural features. This work advances the integration of machine learning with physical modeling for microstructural characterization and fatigue diagnostics.