Learning Vortex Enhancement with Angular-Speed-Invariant Importance Sampling in SPH Fluids

ICLR 2026 Conference Submission14450 Authors

Keywords: Graphics Fluid Simulation, SPH, Vortical Flow, Kinematic Vorticity Number, Learning Fluids

TL;DR: Explore the effect of angular-speed-invariant importance sampling in learning vortex enhancement for SPH.

Abstract: Learning vortex enhancement in SPH benefits most from what is sampled. We use angular-speed-invariant importance sampling with the Kinematic Vorticity Number (KVN) to target vortex cores across resolutions and flow speeds. Particles selected by KVN are pooled into a lightweight global token via attention. Models are trained with velocity correction targets obtained by applying a Biot–Savart mapping to the vorticity loss field. Compared with uniform and vorticity-based sampling, KVN-based sampling improves vortex coherence and advances the emergence of secondary vortices across scenes and particle counts. The gains persist under coarse and fine discretizations and scale smoothly with particle count, indicating robustness to resolution changes. Ablations further show that injecting KVN-based information also benefits alternative encoder variants, suggesting that angular-speed-invariant sampling is a simple, transferable lever for learning vortex enhancement in SPH.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 14450