Curriculum Learning for Mesh-based simulations
This addresses the problem of expensive training for mesh-based simulations in computational fluid dynamics, offering a practical speed-up for researchers and engineers, though it is incremental as it adapts an existing training strategy without modifying the model architecture.
The paper tackles the high computational cost of training graph neural networks on high-resolution unstructured meshes for computational fluid dynamics by proposing a coarse-to-fine curriculum learning approach, achieving comparable accuracy while reducing wall-clock time by up to 50%.
Graph neural networks (GNNs) have emerged as powerful surrogates for mesh-based computational fluid dynamics (CFD), but training them on high-resolution unstructured meshes with hundreds of thousands of nodes remains prohibitively expensive. We study a \emph{coarse-to-fine curriculum} that accelerates convergence by first training on very coarse meshes and then progressively introducing medium and high resolutions (up to \(3\times10^5\) nodes). Unlike multiscale GNN architectures, the model itself is unchanged; only the fidelity of the training data varies over time. We achieve comparable generalization accuracy while reducing total wall-clock time by up to 50\%. Furthermore, on datasets where our model lacks the capacity to learn the underlying physics, using curriculum learning enables it to break through plateaus.