All publications

Featured research · AB-GATr / 2026

When Does Equivariance Help? Canonical Alignment in Neural Fluid Surrogates

Patryk Rygiel, Julian Suk, Kak Khee Yeung, Christoph Brune, Jelmer M. Wolterink

arXiv · Preprint

Earlier title: Symmetry in the Wild: The Role of Equivariance in Neural Fluid Surrogates.

Can a model learn the physics of blood flow without relearning it for every orientation? AB-GATr explores how geometric symmetry and the way data is aligned affect neural fluid surrogates.

CFD reference and model predictions for a rotated vessel, comparing the equivariant AB-GATr with the non-equivariant AB-UPT.
Predictions under changes in vessel orientation. The comparison illustrates the consistency that explicit equivariance brings to this example. Figure 1, Rygiel et al., arXiv v2, CC BY-NC-SA 4.0. Select any figure to view it at full resolution.

Interactive examples

Compare velocity predictions

Why geometry matters

CFD training data is costly, making physical symmetries useful priors for neural surrogates. The incompressible Navier–Stokes operator respects E(3) symmetry—translations, rotations, and reflections—when external forces are absent or transform consistently. Velocity vectors rotate or reflect with the domain; pressure values remain unchanged at corresponding points.

The solution map inherits this symmetry only when geometry, boundary conditions, and initial conditions transform together. Rotating a geometry while keeping its inflow direction fixed changes the physical problem.

In automotive aerodynamics, a wind-tunnel setup fixes the incoming flow direction, making relative orientation physically meaningful. Vascular flows often lack this external reference, although canonical alignment can still provide statistical cues. We study how these differences affect the value of equivariance in Sections 2.1–2.4 of the paper.

The Anchored-Branched Geometric Algebra Transformer

AB-GATr combines sparse geometric anchors with connected branches for surface and volume quantities. Its geometric algebra representation preserves E(3) equivariance while the branched architecture handles different prediction targets, such as velocity within a vessel and shear stress at its wall.

AB-GATr architecture: surface and volume inputs, a shared attention body, and separate velocity and wall-shear-stress decoders.
The encoder, attention body, and branched decoders. Figure 2, Rygiel et al., arXiv v2, CC BY-NC-SA 4.0.

What we found

Symmetry helps blood-flow prediction

Explicit equivariance improves performance across the evaluated hemodynamic benchmarks.

Alignment changes the picture

A controlled automotive experiment shows that strong canonical alignment can favor non-equivariant models.

Augmentation is not equivalent

Rotational augmentation does not consistently reproduce the benefits of an equivariant architecture.

Reference and predicted velocity fields for SynthBifCor, ANEUMO, and AAA-100 vascular geometries, with color scales and example-specific relative errors.
Example predictions across three vascular datasets; the AAA-100 example shows time-averaged flow. The displayed errors describe these examples, not aggregate benchmark scores. Figure 4, Rygiel et al., arXiv v2, CC BY-NC-SA 4.0.

Scope and reproducibility

The conclusions depend on the evaluated tasks and orientation distributions. They do not establish that equivariance is universally beneficial. The paper reports the experiments and limitations in detail.

The official implementation contains the model and instructions for running the experiments. For the vascular geometry dataset, see AAA-100.

Manuscript history

Cite this work

@misc{abgatr2026,
  title = {When Does Equivariance Help? Canonical Alignment in Neural Fluid Surrogates},
  author = { Patryk Rygiel and Julian Suk and Kak Khee Yeung and Christoph Brune and Jelmer M. Wolterink },
  year = {2026},
  doi = {10.48550/arXiv.2605.18816},
  eprint = {2605.18816},
  archivePrefix = {arXiv},
  url = { https://doi.org/10.48550/arXiv.2605.18816 }
}