Physics-informed GCN-LSTM framework for long-term forecasting of 2D and 3D microstructure evolution

This paper presents a physics-informed framework that integrates graph convolutional networks (GCN) with long short-term memory (LSTM) architecture to forecast microstructure evolution over long time horizons in 2D and 3D with remarkable performance. The model compresses phase-field simulation data using convolutional autoencoders and performs prediction in latent graph space. Therefore, it significantly reduces training time, especially in 3D, while maintaining physical fidelity. Physics-based loss terms derived from the Cahn–Hilliard equation, including a mass conservation constraint, are incorporated to improve long-term stability and accuracy. The framework is trained and tested jointly on datasets spanning nine different alloy compositions, and generalizes robustly to an unseen dataset from a new seed without retraining. Long-horizon forecasting evaluations demonstrate strong agreement with ground truth phase-field simulations across different spatial and temporal regimes. This integration of physics-informed learning with graph-based latent dynamics enables efficient and accurate forecasting of microstructure evolution across long temporal and spatial scales.