Acceleration of an algebraic multigrid pressure solver using graph neural networks
Researchers have developed a novel data-driven smoother for algebraic multigrid (AMG) pressure solvers, utilizing a modified graph convolutional isomorphism network (GCIN). This graph neural network predicts optimal polynomial coefficients to construct a sparse pseudo-inverse operator, effectively capturing the algebraic structure of the system and adapting to local anisotropies in unstructured grids. The approach demonstrated significant performance improvements, reducing V-cycles and achieving wall-clock speedups of 4% to 37% across various benchmarks. Notably, the model showed robust generalization capabilities, maintaining efficiency on meshes significantly larger than those used in training and accelerating convergence on industry-relevant problems. AI
IMPACT This research could lead to faster and more efficient computational fluid dynamics simulations by improving the performance of pressure solvers.