Workshops Archive

A Distribution-Function GNN Surrogate for Lattice Boltzmann Flows: Adaptive Coarsening Is Where the Speedup Comes From


Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications

Authors: Hiroki Kaneko and Takashi Shimokawabe (The University of Tokyo)

Abstract: We evaluate a graph-network surrogate for two-dimensional lattice Boltzmann(LBM) flows on both accuracy and wall-clock time per case. The model predicts the increment of the velocity distribution function and reconstructs velocity and density from the LBM definitions; a 49-step rollout (9800 solver steps) completes without diverging. Adding the distribution function to the inputs cuts the velocity error by 20-46%. Static coarsening by variable-radius Poisson-disk selection from geometry alone cuts the node count 22.6-fold while accuracy on the evaluation grid stays on par with the uniform mesh, and cuts the time per case tenfold to 8.9 times faster than the reference solver. On the uniform mesh it is only on par, so the advantage appears solely under coarsening. Against a coarse-grid solver at the same obstacle Reynolds number the surrogate offers robustness, not speed: it returns a solution on all 48 cases, the solver on 42.


Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing

Back to Full Workshop Archive Listing