Overview
This work introduces an AE-GCN-LSTM surrogate framework designed for long-horizon forecasting of microstructure evolution. The framework addresses the computational challenges associated with phase-field modeling in multicomponent and multiphase systems over extended spatial and temporal scales. Specifically, it was applied to the multicomponent AlCrFeNi high-entropy alloy system, which contains coexisting BCC and FCC phases. The method aims to provide a scalable and computationally efficient alternative for simulating microstructure evolution.
Research Context
Phase-field modeling is recognized for its ability to predict microstructure evolution. However, its application becomes computationally demanding for systems characterized by multiple components and phases, especially when simulating over large spatial and temporal scales. The complexity arises from the need to track numerous variables and their interactions over extensive durations, which can render conventional simulations computationally prohibitive.
Approach
The AE-GCN-LSTM framework incorporates a multi-head autoencoder. This autoencoder's function is to compress the four elemental concentration fields and the phase-field order parameter into latent representations. These latent representations are subsequently formulated as graphs to facilitate the learning of their spatial and temporal evolution. The framework was trained using computational domains of $100 \times 100$ size, containing a single nominal alloy composition.
Findings
The AE-GCN-LSTM framework demonstrated accurate forecasting of microstructure evolution over horizons extending to 3,000,000 simulation timesteps. Its robustness was assessed under previously unseen conditions without requiring retraining, fine-tuning, or parameter adaptation. These evaluations included:
- Variations in FCC precipitate size and initial position.
- Microstructures containing one, two, and five FCC precipitates.
- Complex phase interactions, including precipitate merging and splitting.
Although trained exclusively on $100 \times 100$ computational domains with a single nominal alloy composition, the framework was successfully transferred to larger systems, specifically $256 \times 256$ and $512 \times 512$ domains. It also showed successful application to previously unseen AlCrFeNi compositions. Across the evaluated configurations, the model preserved the dominant phase morphology and compositional evolution. The computational speedups achieved ranged from approximately 7200 to 62300 times relative to conventional phase-field simulations.
Why This Matters
The demonstrated scalability and computational efficiency of the latent graph-based AE-GCN-LSTM forecasting method suggest a promising foundation for high-throughput alloy design. Its ability to accurately forecast microstructure evolution over long horizons and generalize to unseen conditions and larger systems could mitigate computational bottlenecks in material science research and development.
Potential Applications
The framework's capability for accelerated simulation of multicomponent, multiphase microstructures offers a foundation for high-throughput alloy design. This suggests its utility in efficiently exploring and optimizing alloy compositions and microstructures.