- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes CEGINO, a lightweight surrogate model that predicts physical fields such as pressure on 3D surface meshes. It builds on EqGINO, which runs Fourier layers on a 3D grid and is therefore only exactly rotation-equivariant for 90°-type rotations.
CEGINO removes the grid. It describes each surface point with rotation-invariant features (distances and projections onto the flow direction), pools them onto ~512 sampled anchor points, mixes the anchors with one message-passing step and one attention layer, and decodes back to the surface.
In a 3-seed comparison on AhmedBody pressure, EqGINO is more accurate in the standard orientation (relative L2 0.19 vs. 0.43). Under unseen arbitrary rotations, CEGINO’s error stays the same, while EqGINO’s rises to 0.88. CEGINO also uses fewer parameters, less memory, and slightly less runtime. Additional single-seed results are reported on ShapeNetCar and DeepJEB.
Critical strengths
-
Rotation-invariant input design. Encoding each point only with invariant quantities (distances and projections onto the flow direction) gives robustness to arbitrary rotations without data augmentation. The matched experiment supports this: CEGINO’s error is essentially unchanged under unseen non-grid rotations, while EqGINO’s more than quadruples. This is the paper’s main idea and its clearest positive result, and it’s the strongest reason to see value in the work.
-
Transparent, controlled comparison. The EqGINO comparison uses a shared split, three seeds, matched hardware, and seed-level statistics. The authors are candid that EqGINO wins on canonical accuracy, and they clearly separate quoted, archived, and new results. This makes the reported trade-off trustworthy, even though it’s narrow in scope.
Critical weaknesses
-
Incremental architecture with a large accuracy cost. The method replaces the grid and FNO with standard components: invariant features, point-cloud pooling, one message-passing step, and a single attention layer over 512 anchors. With only one all-to-all layer, and fine geometry compressed into 512 anchors once with no re-projection, the backbone is weak. This likely explains why CEGINO is about 2.2× worse than EqGINO in the canonical frame, and worse than plain GINO on most AhmedBody targets. The paper doesn’t separate whether the gap comes from the backbone or from the invariant features. As an architecture paper, the novelty doesn’t justify this accuracy loss.
-
The robustness claim rests on a weak baseline and narrow evidence. The rotation advantage is shown only against EqGINO trained without rotation augmentation, a setting where its failure is expected. There’s no comparison with augmented baselines or simple frame canonicalization (aligning the inlet direction), either of which might close the gap at no cost. The evidence also covers a single dataset and target (AhmedBody pressure) with three seeds and high CEGINO variance (±0.119). Since rotation robustness is the paper’s main selling point, this weakens the core claim.
Questions for authors
Clarifications (could change my assessment):
- Neighborhood sizes and mesh density. What k or radius is used for node-to-anchor lifting and for both decoder scales, and what is the “interpolation scale”? Training uses 4× subsampled meshes while testing uses full resolution, so fixed-k neighborhoods span different physical areas. How sensitive is accuracy to this?
- Table consistency. Why are CEGINO’s Table 9 (rotated) values identical to Table 8 (canonical) despite a measured rotation defect of about 0.04? Why does archived CEGINO pressure read 0.3542 in Table 8 but 0.4189 in Table 16?
- Seed variance. Please report per-seed results. What causes CEGINO’s large spread (±0.119 vs. ±0.001 for EqGINO)?
- Practical relevance. Automotive CFD normally uses a fixed frame with a ground plane. In what applications do arbitrary joint rotations actually occur, and is the ground plane encoded?
Presentation and correctness (major problems): 5. Figure 1 does not match the method. It shows Fourier transforms, spectral grids, orbit-based weights, and ODE dynamics, and the caption then says none of these are implemented. Please redraw it to show only the actual pipeline, ideally a layer-level diagram with tensor shapes (N nodes → M anchors → queries). 6. Notation is hard to follow. There are about 11 separate MLP symbols (ψ's). The condition is written inconsistently (c, cᵢ, condᵢ). v is reused for both velocity and displacement. Physics equations (Eqs. 25–27) are included that the method never uses. Meanwhile, key hyperparameters (neighborhood sizes) are missing. Please simplify the notation and define every quantity needed for reproduction. 7. Unjustified “continuous” and iterative framing. The ODE formulation and weight-shared residual dynamics motivate the name, but validation selects L = 1: a single step, with no sharing or continuous depth. Removing anchor dynamics even improves two of eight targets (Table 11). Please justify this design with evidence, or describe it simply as one message-passing layer and revise the title accordingly.
Additional experiments (optional but would most strengthen the paper): 8. Compare against EqGINO or GINO trained with rotation augmentation, and against a simple canonicalization baseline (rotating each sample so the inlet direction is fixed). 9. Run an ablation separating the two possible causes of the canonical accuracy gap: invariant features vs. the weak backbone. For example, use the same invariant inputs with more attention layers or a Transolver-style backbone.
Future work (not required for my assessment): 10. Replace the fixed distance-penalty decoder with learned invariant cross-attention.
3 comments
Q1 — Does a fixed neighbor count cover a larger region on a sparse input? Yes, particularly at lifting. To measure this, we kept the checkpoints, query nodes, labels, sample condition and full-mesh descriptors fixed. We thinned only the nodes available for lifting and anchor selection, using nested subsets of each mesh.
| Input density | Canonical pressure error | Distance to 128th lifting node | Distance to 16th decoder anchor | Distance to 48th decoder anchor |
|---|---|---|---|---|
| Full | 0.3409 ± 0.0320 | 0.0408 | 0.1657 | 0.2904 |
| Half | 0.2928 ± 0.0183 | 0.0581 | 0.1656 | 0.2904 |
| Quarter | 0.2717 ± 0.0153 | 0.0820 | 0.1656 | 0.2903 |
Distances are in normalized mesh units, with equal weight per mesh. At quarter density, the lifting neighborhood reached roughly twice as far, whereas the decoder neighborhoods changed little. Pressure error decreased in this fixed-checkpoint test. Because thinning also changes anchor placement, this does not establish that coarser meshes generally improve predictions or that the trend would persist after remeshing or retraining.
Q2–Q3 — Table provenance and seed variation. Thank you for flagging the identical CEGINO entries in Tables 8 and 9. They are identical at the precision recorded in the archive, but the archive does not let us verify them as two independent measurements. We therefore do not use those entries as evidence of exact equivariance. Equal error norms can coexist with a nonzero equivariance defect: the norms compare each prediction to ground truth, whereas the defect compares the two predictions to one another after accounting for the rotation.
The pressure value 0.3542 in Table 8 is the archived benchmark entry. The 0.4189 in Table 16 comes from a separate, full-resolution evaluation of a historical 1.071M-parameter checkpoint on 50 meshes. The records are insufficient to attribute their difference to a particular setting. Neither value is part of the matched three-seed comparison.
| Canonical pressure relative L2, by training seed | Seed 0 | Seed 1 | Seed 2 |
|---|---|---|---|
| Submitted matched study: CEGINO | 0.2910 | 0.5130 | 0.4763 |
| Submitted matched study: EqGINO | 0.1912 | 0.1889 | 0.1914 |
| New 100-epoch canonical control: CEGINO | 0.3047 | 0.3651 | 0.3528 |
| New 100-epoch canonical control: EqGINO | 0.1907 | 0.1886 | 0.1881 |
The submitted CEGINO runs for seeds 1 and 2 had higher error on the same test meshes. The spread was smaller in the new controls, which selected checkpoints by canonical validation loss. Q4 — Ground plane and practical relevance. Your ground-plane example identifies the use case precisely. Consider a CFD solution exported in two CAD coordinate frames: the geometry and inlet rotate together, but the physical pressure at each corresponding surface point is unchanged. CEGINO handles this change directly, without requiring the two exports to share a canonical orientation. This is distinct from physically rotating a vehicle relative to a fixed road. The model has no explicit ground-plane normal or road boundary input, and we did not vary ride height or forcing relative to the ground. If deployment always uses a fixed frame, EqGINO had the lower canonical pressure error in our comparison.
Q8 — Rotation augmentation and inlet alignment. Following your suggestion, we trained CEGINO, EqGINO and Transolver on the same split with the same Haar-uniform rotation for each seed, epoch and mesh. The clearest remaining advantage for CEGINO was cross-frame consistency: non-grid pressure defects were 0.0387, 0.2057 and 0.1135, respectively. Accuracy gives a different ranking: pressure errors against ground truth were 0.3448, 0.2264 and 0.4337, so augmented EqGINO was most accurate and CEGINO outperformed Transolver. The distinction also held for wall shear stress on the same dataset. CEGINO’s vectors followed the coordinate rotation more closely (non-grid vector defect 0.0309 versus 0.1636), while EqGINO’s vectors were closer to ground truth (three-seed vector error 0.1964 versus 0.2630 for CEGINO).
The inlet-alignment baseline was especially informative. Before applying the frozen, canonical-trained EqGINO, we aligned each observed inlet to +x using the shortest proper rotation about the mesh centroid. This uses neither a reference mesh nor the inverse test rotation, ground normal or target pressure. EqGINO’s non-grid error fell from 0.8843 to 0.2537 ± 0.0019, below CEGINO’s 0.3402 without alignment. The breakdown explains where alignment helps: error was 0.1892 ± 0.0014 on the five z-axis rotations, but 0.4150 ± 0.0101 on the two oblique-axis rotations. An inlet direction suffices to undo the former rotations in ideal arithmetic, but in general leaves rotation about the inlet unresolved. We agree that alignment is a strong baseline when a canonical inlet convention is available. These values are means ± sample SD over three seeds on the specified rotations, not averages over all of SO(3).
Q9–Q10 — Invariant inputs versus backbone capacity. Your proposed ablation separates two explanations that the original comparison could not distinguish. We first varied the inputs while keeping the architecture, input width, initialization and parameter count matched within each seed. Replacing three zero-padding channels with centered Cartesian coordinates changed canonical test error from 0.3223 to 0.3392 and non-grid error from 0.3215 to 0.7005 over three seeds. Restoring pose-dependent information in this way did not recover canonical accuracy and reduced transfer to new orientations.
We then held the invariant inputs fixed and varied capacity. With 512, 1024 and 2048 anchors, canonical validation errors were 0.3379, 0.3477 and 0.3164; four tied local updates at 512 anchors gave 0.3352. To test your attention-depth suggestion directly, we kept 512 anchors, one local update and the training protocol fixed, and trained one- and two-block models for 100 epochs each (seed 0). Their best canonical validation errors were 0.3017 and 0.2939, with 1,203,977 and 1,336,459 parameters, respectively. Together, these tests point to a useful design direction: retain the invariant inputs that supported cross-frame transfer, and improve canonical accuracy by adjusting the anchor budget or mixer capacity. The attention-depth result is a single-seed validation gain, not a measured cross-frame gain for the deeper model. Your suggestion of learned invariant cross-attention at the decoder addresses a separate question from latent attention depth.
We appreciate the concrete alternatives you proposed. They sharpen the practical conclusion: when a canonical frame is unavailable and rotated training examples are not used, CEGINO transfers directly with lower pressure error than unaligned EqGINO on the tested rotations. When alignment or augmentation is available, those are strong accuracy baselines; under matched augmentation, CEGINO’s distinct benefit is that both its scalar pressure and its vector wall shear stress follow a change of frame more consistently.
Thank you for recognizing the value of rotation-invariant inputs and the distinction between our controlled and archived comparisons. Your suggestions on inlet alignment and attention depth helped us test two practical questions: when can preprocessing recover cross-frame accuracy, and can a stronger backbone improve accuracy without giving up invariant features?
Where CEGINO has an advantage. CEGINO can predict pressure in a new coordinate frame without training on rotated examples or first aligning the geometry to a canonical frame. In our new three-seed AhmedBody study, both models were trained only in the canonical frame. Across seven unseen non-grid rotations, CEGINO’s mean relative L2 error against ground truth was 0.3402, versus 0.8843 for unaligned EqGINO. CEGINO’s error was nearly unchanged from its canonical value of 0.3409, whereas EqGINO was more accurate in the canonical frame (0.1892). Thus the benefit is not simply a flatter error curve: CEGINO also gave more accurate predictions in the tested new frames under this training protocol.
Q1, Q5–Q7 — Architecture, notation and the meaning of “continuous.” The architectural choice behind this result is to propagate invariant scalar features on sparse surface anchors rather than on a Cartesian latent grid. In the original matched pressure profile, CEGINO used 1.204M parameters and 1.113 GB of peak GPU allocation, versus 4.154M and 4.293 GB for EqGINO. The table gives the forward pass used in that comparison (batch dimension omitted).
Lifting and decoding use fixed neighbor counts rather than fixed physical radii. The two decoder branches have separate weights: each subtracts normalized squared distance from a learned edge score, then applies a softmax over neighbors. The recorded interpolation scale of 0.05 is an unused compatibility argument, not a kernel bandwidth.
The notation obscured a simpler distinction: the sample condition and unit inlet direction are inputs, while pressure is the predicted field. The governing equations describe the simulations that produced the data; we do not minimize a PDE residual. “Continuous” refers to rotations beyond a finite grid-symmetry group, not continuous network depth or exact numerical equivariance. The selected pressure model takes one residual step, with no adaptive ODE solver. Removing propagation improved some archived targets, so those results do not support a general benefit from repeated steps. Grid-free anchor operators for cross-frame surface-field prediction is a more direct description of what we evaluated.