Bridging Phase-Field Model and Deep Learning for Predicting 2D and 3D Microstructure Evolution in Ternary Alloys
We develop a hybrid framework that integrates a phase-field model (PFM) with an attention-enhanced deep learning (DL) architecture to study ternary spinodal dealloying, a sophisticated self-organization approach used to fabricate three-dimensional bicontinuous, hierarchical nanoporous materials. The study captures three distinct phase-separation mechanisms that emerge during the early stages of spinodal decomposition in both two and three dimensions. The DL workflow consists of three key components: (i) a dimensionality-reducing autoencoder that provides compact representations of high-resolution microstructure images (256x256x3), (ii) an attention-augmented convolutional long short-term memory (ConvLSTM) network that learns complex spatiotemporal correlations governing microstructure evolution, and (iii) a novel slice-by-slice strategy that enables extension of the model to three-dimensional systems (128x128x128x3). We further demonstrate a hybrid simulation strategy in which PFM accurately captures rapid early-stage microstructure evolution, while the DL model efficiently predicts late-stage coarsening dynamics. The trained DL model achieves remarkable predictive accuracy, maintaining fidelity up to 400 timesteps ahead and generalizing to compositions outside the training distribution. By bridging the physical fidelity of PFM with the computational efficiency of DL, this framework establishes a robust platform for predictive modeling of microstructure evolution in complex multicomponent systems.