Training CFD Surrogate Models: Why Sampling Strategies Differ from LLMs and Images

capetorch · x · 2026-08-12

The author explains that sampling strategies for training CFD surrogate models differ significantly from modalities like LLMs and images. Since each sample is a massive point cloud with millions of points, efficient training typically requires a two-level dataloader: sampling a case first, then dynamically subsampling its surface and volume points. This asymmetry also makes the concept of an epoch ill-defined.

Related event: Efficient 3D Data Sampling Strategies for CFD Surrogate Models(2 posts)→

Original post →

More from Research

Research channel →