A PDE predictor stays steady when its description changes
ISLA: A Coordinate-Free, Measure-Weighted Transformer for Boundary-Driven PDEs · arXiv:2611.17141
In plain words
A physical equation can describe one problem in different coordinate frames, units, or boundary meshes. A predictor should still give the same physical answer, but some models change their predictions when those descriptions change.
ISLA is a neural method for predicting fields from boundary information. It weights boundary cells by the area they represent and lets interior points read from the boundary. In tests that changed a problem’s pose or boundary sampling, ISLA’s prediction errors changed little across the tested tasks.
The research
A prediction that shifts when axes, units, or mesh sampling change can make it harder to reuse a surrogate for the same physical problem.
ISLA’s design and tests address that specific reliability question: whether the predicted field stays stable across selected changes in how a problem is represented.
ISLA is a neural surrogate for steady boundary-value equations that encodes geometric relations and boundary area so predictions can ignore arbitrary presentation choices.
ISLA encodes geometric relations, scales lengths using the boundary, weights boundary-cell contributions by represented area, and lets interior queries read without contributing to boundary sums.
The authors prove presentation-related properties under stated assumptions and mesh-refinement convergence under separate conditions.
Tests that changed pose or boundary sampling found little change in ISLA’s errors across the tested tasks, while coordinate-based baselines could worsen sharply.
On AeroZoo’s full test split, ISLA also had the lowest reported mean reference-scaled velocity and pressure errors among the baseline settings.
The formal guarantees rely on stated conditions such as exact arithmetic and, for some interior claims, an exact boundary read. The empirical tests cover selected flow and mechanics tasks; presentation independence did not ensure accurate predictions for every unseen boundary layout.
What the evidence shows
- The presentation problem
Invariant Slice Attention (ISLA) is a neural surrogate method for predicting fields in steady boundary-value equations. Its central claim is that predictions should not change just because the same physical problem is expressed with different axes, units, boundary meshes, or requested interior points.
That matters because coordinate-based surrogates and unweighted point sums can make these arbitrary presentation choices affect the answer.
- What ISLA changes
ISLA changes how geometry and boundary cells enter the model: it uses relations to learned anchors rather than raw learned coordinate inputs, weights boundary contributions by represented area, and makes interior queries read-only.
The authors’ mathematical results establish presentation-related properties under explicit assumptions, including independence between requested interior points. They also prove mesh-refinement convergence under separate conditions.
- Testing the same problem
These tests compare each model’s error on the same cases before and after changing the pose.
They show that ISLA’s errors stayed nearly fixed across the tested tasks, while tuned coordinate-based models could become much less stable on DrivAerML pressure; this measures presentation stability, not overall accuracy.
- Changing the cell sample
Changing boundary sampling also tests whether a model depends on the mesher’s pattern. On three surface tasks, ISLA’s errors stayed nearly unchanged when cells were sampled in proportion to area.
For DrivAerML pressure, the resampled-to-stored error ratio was 7.2 for Transolver and 3.6 for GeoTransolver. The shared scoring weights make the comparison about the presentation change, while the result remains limited to those tested surfaces.
- A broader flow test
AeroZoo adds a different test: predicting flow across cases with varied geometry, scale, speed, pose, and boundary conditions. On its 515-case full test split, ISLA had a lower mean reference-scaled velocity error than every reported baseline setting.
This supports the method’s predictive performance on that dataset as well as its stability under selected presentation changes, but does not imply accurate extrapolation to every unseen boundary layout.
Deep
- Original paper authors
- Peter Sharpe
- Submitted to arXiv
- License
- CC BY 4.0