Graph model simulates grain growth faster than phase-field calculations

The method, called VN-GraphGG, converts a two-dimensional polycrystalline microstructure into a graph with grain and junction nodes. At each step, grain areas are updated by a physical law tied to the number of sides, then a graph neural network reconstructs the next microstructure field. Separating these stages allows kinetic coefficients to change without retraining the reconstruction component.

On a 6,000-example test set, the authors report a mean pixel-classification error of 3.54% and a mean grain-area error of 2.04 pixels per grain. They also report similar bulk coarsening behaviour to phase-field simulations at two orders of magnitude lower computational cost. These are simulation benchmarks; the paper does not report experimental validation in a real production process.