SimGate: A deep learning surrogate model for predicting microstructure evolution using the phase-field method

Abstract

This study introduces SimGate, a novel deep learning surrogate model for predicting microstructure evolution using the phase-field method. Combining the temporal modeling capabilities of “Simpler yet better video prediction (SimVP)” with the multi-order aggregation features of “Multi-order gated aggregation network (MogaNet)”, SimGate leverages robust temporal dynamics alongside spatial and channel aggregation modules to ensure precise detail capture and spatial consistency. To demonstrate SimGate’s ability to tackle challenging scenarios, high-temperature sintering simulations of polycrystalline cerium dioxide (CeO₂) particles were selected as a test case. These simulations, chosen for their complexity, involve both Cahn–Hilliard-type and Allen–Cahn-type phase-field equations along with intricate interfacial dynamics, and they were validated through experimental data. SimGate accurately predicts the sintering process from limited initial time steps and exhibits strong extrapolation capabilities in modeling unseen microstructures over extended time scales. Compared to traditional phase-field simulations, which require hours per case, SimGate reduces computational time to seconds while maintaining a prediction accuracy of around 90%. Additionally, point-wise error analysis shows that the average accuracy is improved by 7.80% and 12.41% compared with the original SimVP and well-known Long Short-Term Memory Networks (LSTM), respectively. An ablation analysis was performed to reveal the contributions of key components in the proposed SimGate framework. By significantly enhancing computational efficiency and accuracy, SimGate demonstrates broad potential as a generalizable microstructure prediction model applicable to diverse material and mechanical processing scenarios beyond sintering.