acceptodds
Under review as a conference paper at ICLR 2027

Coordinates Are Labels, Not Physics: Breaking the Coordinate Shortcut in Neural Surrogates

Abstract

Neural surrogates approximate physical fields at a lower computational cost than numerical simulation. On aligned vehicle benchmarks, a shared pose and fixed inflow direction allow models to predict field values from absolute coordinates, a strategy we call the coordinate shortcut. Low error on these benchmarks therefore does not show whether a model has learned how the field responds to geometry and boundary conditions. We evaluate surrogates beyond this setting, under unseen inflow directions and on geometries from unrelated shape families. Removing coordinate inputs prevents the shortcut but also discards spatial information about the arrangement of points within a geometry. We propose GraFINS (Graph-Encoded Frame-Invariant Neural Surrogate), which replaces coordinate inputs with frame-invariant features encoding local geometry and its relation to boundary conditions, and a graph embedding encoding spatial information. Trained only on vehicles with flow from the front, GraFINS predicts pressure under side and rear inflow more accurately than equivariant and non-equivariant baselines given the same features, and without retraining it also outperforms them on non-vehicle geometries. It preserves in-distribution accuracy on automotive and patient-specific blood-vessel benchmarks and remains stable under scene rotations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operator, Surrogate Model, Physics AI

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_Exb9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies reliance on dataset-frame coordinates in neural CFD surrogates trained on aligned geometries. GraFINS replaces direct coordinate inputs with boundary-aware invariant features and a multiscale anchor-graph embedding, using Transolver or Transolver-3 as the backbone. Evaluation covers automotive and vascular benchmarks, separately simulated inflow directions, unrelated synthetic shape families, and joint rotations of geometry and boundary conditions.

Critical strengths
  • The generalization experiments address a meaningful limitation of current benchmarks. The unseen-inflow study changes the boundary condition on a fixed geometry and recomputes the CFD solution, while the unrelated-shape study evaluates frozen checkpoints. GraFINS achieves a mean R2 of 0.480 on unseen inflows versus 0.338 for AB-GATr, and 0.356 on unrelated shapes versus 0.024 for feature-only Transolver (Table 3). The unseen-inflow advantage also holds with a single inference pass.

  • The graph provides useful spatial information without sacrificing much in-distribution accuracy. Table 2 shows that it restores the accuracy lost when coordinates are removed, with evidence across two backbones and multiple prediction targets. On Aneumo, for example, relative L2 error decreases from 46.34% for Transolver with coordinates and invariant features to 36.39% for GraFINS. The spatial probes and graph ablations provide additional support for this representation.

Critical weaknesses

A direct alternative to the proposed coordinate replacement is missing. In App. C.1, I see the paper already constructs an orthonormal frame from the inflow and ground directions, and the proposed features include streamwise position and height (Table 7). Feeding coordinates expressed in this physical frame to the same backbone would retain spatial information while removing dependence on dataset axes. This comparison is particularly relevant because invariant representations and alignment sensitivity are already established in the cited literature (App. A.1). Without it, the methodological advantage over a suitable coordinate convention remains unclear.

The main ablation does not isolate coordinate removal from the added graph model. GraFINS introduces a learned spatial encoder, increasing the parameter count from 5.96M to 8.87M (Table 9), but no otherwise identical graph-equipped model retains coordinates. The spatial probes and the AhmedML set-attention comparison support the graph’s utility; they do not establish whether excluding coordinates is responsible for the main generalization gains. This missing control also leaves the categorical claim in Sec. 4.2—that responding to unseen inflow requires inputs untied to the dataset frame—insufficiently supported.

Questions for authors

How does Transolver perform on the Table 3 tasks when its coordinate inputs are expressed in the inflow–ground frame already constructed in App. C.1, alongside the same invariant features?

How does GraFINS perform on those tasks when coordinates are retained, with the graph, prediction head, training recipe, and inference settings otherwise unchanged? This would directly test the contribution of coordinate removal.

Reviewer_QqhJOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper advocates for coordinate-free representation learning to bypass shortcuts learned by neural surrogate models. GraFINS is proposed to center the graph representation learning on frame-invariant features (eq1 and App. C.1) with multi-scale anchors and a Transolver-based coordinate-free backbone. Results demonstrate promising generalization for unseen geometry, inflow direction, and coordinate frames.

Critical strengths

strengths:

  1. The insight on removing coordinates and use only frame-invariant features for improved generalization is technically novel and interesting with potential to be applicable beyond the tested domains
  2. The experiments are well-prepared: sanity checks on coordinate short cut, and compelling results from GraFINS with multi-faced generalization setups in Sec. 4

Overall, I find the coordinate shortcut insight and the proposal of GraFINS to be technically novel, and superior results have been demonstrated in challenging cases that GraFINS is proposed to fix. This is the core merit of the paper.

Critical weaknesses

weaknesses:

  1. Confounding variable in the superior performance of GraFINS not fully controlled:
    • can be the heavier-weight network compared to Transolver baselines
    • can be multi-scale modeling
  2. generality of handcraft frame-invariant features beyond the evaluated simulations

W1: With GraFINS attaining the best results across different generalization scenarios, despite the frame-invariant feature engineering and learning being the core proposal to solve the coordinate shortcut, some factors that might contribute to the success are not fully isolated.

For example, as Table 9 shows, GraFINS uses more params: ~3M more compared to Transolvers, ~4-8M more compared to equivariant baselines, with longer training and inference time. Therefore, an iso-param and an iso-FLOP controlled experiment would further validate whether the success is largely attributed to additional param or more operations.

A second confounding variable is the multi-scale graph representation learning. This seems to be orthogonal to the frame-invariant features, which might be an important aspect of the success. Ablation studies that isolate frame-invariant features as the input vs. multi-scale modeling would clarify the contribution of each factor.

It would be interesting to hold the feature transformations across Transolver baselines and GraFINS the same: either remove or add proposed multi-scale modeling altogether, and report the result.

W2: While several formulations for the simulations-of-interest are proposed, the frame-invariant feature appears to be highly domain-specific. A strong addition to the paper can be a discussion and demonstration of the promise of frame-invariant features across domains. I understand this requires cross-domain scientific expertise, but perhaps a useful direction is to additionally propose and evaluate more general-purpose frame-invariant features insensitive to domain science, for example, ones derived from coordinates alone.

Questions for authors

Summarizing the weaknesses section:

W1: request for additional experiments to rule out confounding variables:

  1. iso-param and an iso-FLOP-controlled experiment
  2. Ablation studies that isolate frame-invariant features as the input vs. multi-scale modeling

W2: request for strengthening the generality of the frame-invariant modeling idea:

  1. Discuss and evaluate more general-purpose frame-invariant features insensitive to domain science, for example, ones derived from coordinates alone.
Reviewer_vGS7OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the coordinate shortcut in neural surrogates. It proposes GraFINS, which replaces raw coordinate inputs with boundary-referenced frame-invariant features and a multi-scale anchor graph that preserves spatial information. Experiments demonstrate competitive in-distribution accuracy across automotive and vascular benchmarks, together with improved generalization to unseen inflow directions and unrelated geometric shapes. The method also maintains stable predictions under joint scene rotations.

Critical strengths
  1. Meaningful problem formulation and evaluation. The paper identifies an important limitation of conventional aligned CFD benchmarks. The coordinate predictability analysis provides a convincing motivation, while the unseen-inflow experiments test genuine changes in boundary conditions rather than simply rotating the coordinate frame. The arbitrary-geometry experiments further extend the evaluation beyond the training shape distribution.

  2. Effective geometric representation with substantial empirical improvements. The combination of frame-invariant features and multi-scale graph embeddings provides a practical alternative to raw coordinates without sacrificing in-distribution accuracy. GraFINS consistently outperforms the evaluated baselines under unseen inflow and geometry changes, including comparisons with the same Transolver backbone using different inputs. The results across two backbones, multiple datasets, and additional spatial probes provide useful evidence for the effectiveness of the proposed representation.

Critical weaknesses
  1. Further clarification of the generalization mechanism would be helpful. The experiments demonstrate convincing OOD improvements, and the proposed representation appears effective. It would be interesting to better understand the individual roles of coordinate removal and the graph embedding, particularly whether the graph could also capture some dataset-specific positional information (see Questions 1–2).

  2. Room for further improvement in absolute OOD accuracy. While GraFINS achieves substantial gains over the baselines, its absolute accuracy remains moderate on some unseen geometries. This is understandable given the challenging zero-shot setting, but a brief discussion of the remaining limitations would further strengthen the evaluation (see Question 3).

Questions for authors
  1. Could the authors compare (f(\phi,g)) with (f(p,\phi,g)) to isolate the effect of removing absolute coordinates on OOD generalization?

  2. Could the authors evaluate (f(g)) alone to assess whether the graph embedding also captures dataset-specific positional shortcuts?

  3. Could the authors provide further analysis of prediction errors in challenging regions, such as separated or concave flow regions, or evaluate integrated quantities such as pressure-induced forces? This would help assess the physical significance of the reported OOD improvements.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.