← Papers
Transformers

Transolver-3

Transformer solvers on industrial meshes of over 100 million cells

Paper title Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries

Zhou et al. (Tsinghua University) · Aero & CFD · Public: code

Paper arXiv 2602.04940PDF

Keeps Transolver's physics-attention, where mesh cells are softly grouped into a small set of learned physical states that attend to each other, but reorders its matrix products and tiles the geometry into chunks to cut peak memory. Training uses random subsets of each mesh (100 thousand cells per sample on DrivAerML); at inference the physical states are accumulated over the whole mesh into a cache, and every cell is then decoded from that cache.

Published
ICML

Share on LinkedIn, X or Bluesky, or or .

Know related work that belongs here?

Something wrong on this page?

More papers

All papers
  1. TransformersMNC — Retrofitting foundation models to unstructured meshes
  2. TransformersPSP grounding — Correcting a CFD-trained wing surrogate with wind-tunnel pressure paint
  3. TransformersTransolver without attention — Does Transolver really need a Transformer?