acceptodds
Under review as a conference paper at ICLR 2027

CEGINO: CONTINUOUS EQUIVARIANT GEOMETRY- INFORMED NEURAL OPERATORS FOR 3D PDES

Abstract

Neural operators offer fast surrogates for partial differential equations on irregular 3D geometries, but coordinate-dependent representations can make their predictions sensitive to changes in reference frame. EqGINO addresses this issue through equivariant Fourier operators, yet its exact rotational equivariance is restricted to grid-compatible rotations. This limitation motivates representations that can support rotational robustness beyond a discrete symmetry group without relying on a Cartesian latent grid. To this end, we introduce CEGINO, a compact geometry-informed neural operator for surface-field prediction. CEGINO encodes geometry through rotation-invariant descriptors, propagates features over sparse anchors, and reconstructs physical fields through two-scale query decoding. Global context complements local propagation, with invariant anchor attention included in the matched pressure configuration. Scalar outputs use invariant decoding, while wall-shear-stress vectors are reconstructed in a rotation-consistent local basis. We evaluate the framework on AhmedBody, ShapeNetCar, and DeepJEB, and conduct a controlled three-seed comparison with EqGINO on AhmedBody pressure to examine the resulting accuracy–robustness trade-off. EqGINO achieves higher accuracy in canonical and grid-compatible orientations, whereas CEGINO better preserves predictive performance under the tested unseen non-grid rotations without rotation augmentation, while requiring fewer parameters, lower forward-pass latency, and less memory. These findings support sparse invariant representations as a practical approach to balancing rotational robustness and computational efficiency in surface-field surrogate modeling. Code is available at https://anonymous.4open.science/r/CEGINO-9756.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, equivariant learning, scientific machine learning, irregular geometries

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DEjAOpenReview ↗
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
  1. 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.

  2. 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
  1. 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.

  2. 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):

  1. 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?
  2. 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?
  3. Seed variance. Please report per-seed results. What causes CEGINO’s large spread (±0.119 vs. ±0.001 for EqGINO)?
  4. 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

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).

StageOperationOutput
InputCenter the coordinates, scale to unit maximum radius, and form invariant scalar features from the geometry, inlet direction, surface normals and sample conditionN surface points
LiftSelect 512 anchors by farthest-point sampling; pool the 128 nearest input nodes for each anchor512 × 128
PropagateApply one local residual update; radius 0.1 in normalized coordinates, at most 24 neighbors, no self-edges512 × 128
MixApply one four-head self-attention block to the anchor features512 × 128
DecodeAt each query, combine features from its 16 and 48 nearest anchors with pooled context, the sample condition and axial positionQ × 1 pressures

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.

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 densityCanonical pressure errorDistance to 128th lifting nodeDistance to 16th decoder anchorDistance to 48th decoder anchor
Full0.3409 ± 0.03200.04080.16570.2904
Half0.2928 ± 0.01830.05810.16560.2904
Quarter0.2717 ± 0.01530.08200.16560.2903

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 seedSeed 0Seed 1Seed 2
Submitted matched study: CEGINO0.29100.51300.4763
Submitted matched study: EqGINO0.19120.18890.1914
New 100-epoch canonical control: CEGINO0.30470.36510.3528
New 100-epoch canonical control: EqGINO0.19070.18860.1881

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.

Reviewer_g2RZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CEGINO, a neural operator for predicting surface fields on 3D geometries that avoids the Cartesian latent grid used by EqGINO. It uses rotation-invariant point features, aggregates them onto a sparse set of anchors, passes messages between anchors, and decodes the fields back at the surface points. Vector fields are reconstructed in a local basis. In a three-seed comparison on AhmedBody pressure, EqGINO is more accurate in the canonical orientation, while CEGINO keeps a similar error under unseen rotations and uses fewer parameters, less memory, and less time.

Critical strengths

S1. The model is lighter than the main baseline in parameters and memory, which can matter for large 3D surface meshes.

S2. The authors are open about many limitations of their evaluation, and they provide code.

Critical weaknesses

W1. The accuracy is much worse than the main baseline, and the robustness claim is not well supported.

  • W1.1 In the controlled comparison, CEGINO's canonical error is more than twice that of EqGINO, with large variance across seeds. Under non-grid rotations, CEGINO keeps this same, much worse error.

  • W1.2 The robustness comparison is limited to models trained without rotation augmentation. Since CEGINO uses rotation-invariant inputs, its flat error under rotation is expected by design, and this alone does not show that it is a better operator. A more informative comparison would train all models with rotation augmentation, including EqGINO and non-equivariant baselines such as GINO and Transolver. Currently, non-equivariant baselines appear only in archived single-seed results, which the authors say generally favor GINO and EqGINO.

W2. The vector-field reconstruction is unclear and not evaluated. The choice of local basis is not justified, there is no study of how it affects accuracy, and points where the basis degenerates are not discussed. Since the inputs discard chiral information, it is also unclear how the model handles mirror-symmetric shapes. Wall-shear-stress prediction is not part of the controlled comparison.

W3. The efficiency claim and the role of anchors are not analyzed.

  • W3.1 The reported latency excludes preprocessing. Anchor sampling, neighbor search, and graph construction may be a large part of the real cost, so the speed-up may not hold end to end.

  • W3.2 Results are sensitive to the number of anchors, with no clear trend, based on a single seed. The authors also note that anchor selection can affect rotational consistency, but this is not studied.

W4. The presentation does not match the method. Figure 1 shows Fourier transforms, a spectral grid, and orbit-based weights, which the caption says are not in the implementation. "Continuous" refers to an ODE view that is not actually used, and "equivariant" mostly means invariant inputs.

Questions for authors

Q1. (W1) Could you train all models with rotation augmentation, including CEGINO, EqGINO, and non-equivariant baselines (e.g., GINO, Transolver), and compare them under the same protocol?

Q2. (W2) How is the local basis handled at degenerate points? How sensitive is wall-shear-stress accuracy to the choice of basis? Could you add wall shear stress to the controlled comparison?

Q3. (W3) What is the end-to-end runtime, including anchor sampling and graph construction? How sensitive are accuracy and rotational consistency to anchor placement, across multiple seeds?

Q4. (W4) Could you replace Figure 1 with a figure of the actual architecture?

5 comments

W1/Q1 — Accuracy under a change of frame. A flat error curve alone does not establish accuracy. The relevant test is error against ground truth in the same new frame. In fresh canonical-trained runs on AhmedBody pressure, CEGINO had lower relative L2 error than EqGINO on seven unseen non-grid rotations (0.3402 vs. 0.8843), despite higher error in the canonical frame (0.3409 vs. 0.1892). The non-grid improvement is therefore a difference in prediction accuracy, not just consistency.

We also trained a non-equivariant baseline, Transolver, and both primary models with the same Haar-uniform SO(3) rotations, keyed by seed, epoch and mesh and applied before model-specific preprocessing. We retained the 413/45/50 split, 100 epochs and canonical validation selection. Learning rates and training subsampling were model-specific. Below, values are means ± sample SD across three training seeds; each non-grid score first averages the seven rotations on the same 50 meshes within each seed.

TrainingModelCanonical errorNon-grid errorNon-grid defect
CanonicalCEGINO0.3409 ± 0.03200.3402 ± 0.03190.0325 ± 0.0043
CanonicalEqGINO0.1892 ± 0.00140.8843 ± 0.00370.8823 ± 0.0056
Shared augmentationCEGINO0.3457 ± 0.00510.3448 ± 0.00470.0387 ± 0.0088
Shared augmentationEqGINO0.2575 ± 0.00890.2264 ± 0.00910.2057 ± 0.0057
Shared augmentationTransolver0.4225 ± 0.02190.4337 ± 0.02360.1135 ± 0.0295

Augmentation changes the ranking: EqGINO becomes the most accurate, while CEGINO has the smallest frame-consistency defect and lower non-grid error than Transolver. Thus our accuracy advantage over EqGINO is transfer without rotation training, not superiority under every training regime.

W2/Q2 — Vector reconstruction. The WSS head predicts scalar coefficients along the inlet direction and two transverse axes constructed by cross products with the centered radial direction. Under a proper rotation the axes rotate together; the coefficients need not change. At exact alignment the transverse axes vanish: the denominator floor avoids division by zero but cannot recover the missing directions. This is an SO(3) argument, not a claim about reflections or chirality.

Of 3,598,547 vertices on the 50 test meshes, 14 had an inlet–radial angle sine below 0.001, and none fell below 0.0001. For one archived checkpoint, vertices below 0.01 contributed on average 0.01949% of the squared equivariance residual per non-grid mesh–rotation pair (maximum 0.2914%). This locates the observed residual; it does not establish basis insensitivity or a WSS advantage over EqGINO.

We tested WSS directly, scoring the vector, not its magnitude, against the correspondingly rotated ground truth. Both models used the same Haar training rotations, three seeds and 50 test meshes. Across seven non-grid rotations, EqGINO had lower vector error, but CEGINO had a smaller equivariance defect: it followed the change of frame more closely. Accuracy against ground truth and consistency across frames answer different questions.

WSS modelCanonical errorNon-grid errorNon-grid vector defect
CEGINO0.2628 ± 0.04250.2630 ± 0.04240.0309 ± 0.0039
EqGINO0.2290 ± 0.00450.1964 ± 0.00140.1636 ± 0.0023

To test basis choice, we rotated the two transverse basis axes by a position-dependent angle, preserving the singular set and parameter count, and retrained the model. In the completed seed-0 pair (100 epochs each), non-grid vector error changed from 0.2157 to 0.2351, and defect from 0.0270 to 0.0298 (original → alternative basis). Basis choice therefore affected this run.

W3/Q3 — End-to-end cost and anchor placement. We timed the full path from opening a mesh file to returning predictions to the host, including anchor sampling and graph construction, on the same 50 meshes with warm filesystem caches. Each of three checkpoints was timed twice, reversing model order; initialization was excluded. CEGINO’s forward pass was faster (105.2 vs. 117.4 ms/mesh), but preprocessing cost more (22.2 vs. 1.8 ms). End to end, CEGINO took 139.8 ± 7.6 vs. 131.5 ± 9.2 ms/mesh. The demonstrated resource advantage is model size and GPU memory, not end-to-end speed: 1.204M vs. 4.154M parameters and 1.113 vs. 4.293 GB peak allocation in the original profile.

Across three seeds, canonical validation errors at 512, 1024 and 2048 anchors were 0.3379, 0.3477 and 0.3164; this is not a monotonic trend. To isolate placement from capacity, we kept the trained 512-anchor models fixed and changed only the farthest-point sampling start, using corresponding vertices before and after rotation. The alternative starts retained only 1.22% and 1.08% of the original anchor IDs on rotated meshes. Yet the three starts gave non-grid errors of 0.3402, 0.3409 and 0.3378, and defects of 0.0325, 0.0316 and 0.0315. Across the three tested starting vertices, both non-grid prediction error and equivariance defect changed little.

W4/Q4 — Architecture diagram and its correspondence to Section 3. In the current architecture diagram, stages 1–2 show the geometry-aware lifting in Section 3.3: we form rotation-invariant scalar features from the surface geometry and inlet direction, then aggregate them onto 512 anchors selected by farthest-point sampling. Stages 3–4 show the two operations in Section 3.4. In the matched pressure model, one residual message-passing step exchanges features between neighboring anchors; one four-head self-attention block then connects all anchors. Stages 5–6 correspond to the decoder and readout in Section 3.5: each surface query aggregates features from its 16 and 48 nearest anchors to predict scalar pressure. For wall shear stress, the readout instead predicts coefficients in a local basis that rotates with the input.

CEGINO addresses a specific physical requirement: when the same problem is expressed in a rotated coordinate frame, the geometry and inlet direction rotate together. At corresponding surface points, pressure should remain unchanged, while the wall-shear-stress vector should rotate with the frame. The vector head implements this distinction by combining invariant coefficients with a rotating local basis where that basis is non-degenerate. “Continuous” refers to the rotation domain beyond a finite grid-symmetry group, not to an ODE solver. W1 and W2 report how closely the predictions follow these transformation laws in practice.

Reviewer_YN6rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CEGINO, a compact geometry-informed neural operator for surface-field prediction on irregular 3D geometries. The core idea replaces EqGINO's Cartesian Fourier grid with sparse invariant anchors: rotation-invariant geometric descriptors encode the input, features propagate over a sparse anchor graph, and a two-scale query decoder reconstructs physical fields. Scalar outputs use invariant decoding, while wall-shear-stress vectors are reconstructed in a rotation-consistent local basis.

Evaluation covers AhmedBody, ShapeNetCar, and DeepJEB benchmarks. The central controlled comparison (AhmedBody pressure, 3 training seeds) shows that EqGINO achieves higher accuracy in canonical and grid-compatible orientations (Rel-L2: 0.1905 vs 0.4267), but CEGINO better preserves predictive performance under unseen non-grid rotations (error nearly unchanged vs increasing from 0.19 to 0.88), while requiring 71% fewer parameters, 20.5% lower inference latency, and 74.1% less GPU memory.

Critical strengths

S1. Clear problem formulation and rigorous experimental protocol. The paper correctly identifies that EqGINO's exact equivariance is restricted to grid-compatible rotations and designs a sound controlled comparison: identical data splits, matched seeds, same test meshes, zero-shot rotation evaluation. Three-seed paired t-tests, equivariance defect metrics, and separate grid/non-grid analysis demonstrate experimental rigor.

S2. Significant resource efficiency gains. Anchor compression from N surface nodes to M=512 anchors reduces parameters from 4.15M to 1.20M, memory from 4.29GB to 1.11GB, and inference latency from 0.158s to 0.126s/mesh. This accuracy-efficiency trade-off has practical value in resource-constrained deployment.

S3. Complete equivariance theoretical analysis. Appendix C systematically derives rotational invariance from input descriptors through anchor lifting, residual propagation, global mixing, to cross-decoding, explicitly stating assumptions on discrete correspondence and non-degenerate bases. This transparency is commendable.

Critical weaknesses

W1. Excessive canonical accuracy gap CEGINO's canonical Rel-L2 (0.4267) is 2.24× that of EqGINO (0.1905). The paired test yields p=0.076, failing to reach 0.05 significance with 3 seeds. More critically, EqGINO achieves lower error on all 50 test meshes. This gap undermines the practical value of the claimed "accuracy-efficiency trade-off"—in engineering applications, >2× error is typically unacceptable. The paper does not explore whether increasing anchor count, residual depth, or training strategies could narrow this gap.

W2. Missing comparison with rotation-augmented EqGINO The core selling point is zero-shot non-grid rotation robustness, but only canonical-trained EqGINO checkpoints are compared. A natural question: if EqGINO also uses continuous rotation augmentation, would its non-grid robustness improve substantially? Table 10's continuous-rotation results suggest EqGINO benefits from augmentation (pressure Rel-L2: 0.1848 vs CEGINO's 0.4295)—the gap remains significant. No controlled experiment compares both under rotation augmentation.

W3. Information loss from invariant descriptors not quantified Equations (4-5)'s invariant descriptors (flow projection, radial distance, orthogonal distance) discard azimuthal and chiral information. The paper acknowledges in Appendix H that "invariant scalar descriptors can discard azimuthal and chiral information" but does not quantify this loss's impact on prediction accuracy. For asymmetric geometries or flow-insensitive features, does this loss contribute to canonical accuracy degradation?

W4. Figure 1 conceptual diagram inconsistent with implementation Figure 1 shows "Fourier Transform," "Spectral Grid," and "Orbit-based Weights" components, but the main text explicitly states "the archived four-step instance reuses weights. It contains no Fourier transform, spectral convolution, or orbit-based spectral weights; those labels in the conceptual artwork are not implementation claims." This disconnect between illustration and implementation risks misleading readers and should be revised.

Questions for authors

Q1. CEGINO's canonical error is 2.24× EqGINO's. Can the gap be narrowed by increasing anchor count (e.g., 1024 or 2048) or introducing multi-scale anchor hierarchies? Table 3 shows 1024 anchors perform worse—does this indicate saturation near 512, or other bottlenecks?

Q2. Could the authors provide EqGINO's zero-shot rotation results under identical continuous rotation augmentation training? Table 10's continuous-rotation results suggest EqGINO also benefits, but controlled comparison is lacking.

Q3. What is the physical interpretation of equivariance defect? Table 15 shows CEGINO's non-grid defect is 0.0486 vs EqGINO's 0.8772. What is the relationship between defect and prediction error? Does low defect mean predictions are "consistent but inaccurate" under rotation?

Q4. WSS vector prediction's local basis guarantees equivariance under non-degenerate conditions. In practice, how frequently do degenerate or near-degenerate configurations occur (e.g., flow direction parallel to radial direction)? Does the paper analyze the spatial distribution of WSS equivariance defect?

Q5. Could the authors provide per-mesh visualization of CEGINO under different rotation angles? Figure 2's aggregated curves mask mesh-level variability—some meshes may perform poorly under specific rotations.

4 comments

Thank you for identifying the key practical question: if training data are available only in a canonical frame, how well does a model predict the same physical field when the coordinate frame changes? This is where CEGINO shows its clearest advantage. In our fresh AhmedBody pressure comparison, jointly rotating the geometry and inlet direction to seven non-grid orientations left CEGINO’s mean relative error nearly unchanged (0.3409 canonical; 0.3402 rotated). EqGINO was more accurate canonically (0.1892), but its error rose to 0.8843 under the same rotations. CEGINO therefore achieved approximately 62% lower prediction error on these unseen orientations without rotation augmentation. The additional experiments below distinguish this accuracy advantage from rotation consistency alone.

W1/Q1 — Canonical accuracy and capacity. We agree that EqGINO is preferable on accuracy alone when the deployment frame is fixed and canonical. The original paired test (p = 0.076) does not establish equivalent accuracy. CEGINO offers a different trade-off: in the original matched profiling setup, it used 71.0% fewer parameters, 20.5% lower forward-pass latency, and 74.1% less peak allocated memory, while better preserving accuracy when the frame changed.

To investigate whether anchor compression limits canonical accuracy, we repeated every seed-0 capacity variant with two additional seeds. With one propagation step, mean canonical validation error was 0.3379 ± 0.0314 at 512 anchors, 0.3477 ± 0.0446 at 1024, and 0.3164 ± 0.0092 at 2048. Four tied steps at 512 anchors gave 0.3352 ± 0.0102. Thus, the worse 1024-anchor result does not establish saturation at 512. Increasing the anchor count to 2048 lowered mean validation error by about 6.4% without adding trainable parameters, although it increases computation and improved on the control in only two of three seeds. We did not evaluate these variants on the test set or test a multiscale hierarchy, so this is a promising capacity result, not evidence that the canonical test gap has closed.

W2/Q2 — Does the advantage persist under identical rotation augmentation? We trained both models from scratch with and without Haar-uniform SO(3) augmentation, using three seeds per condition (12 runs). For each seed, epoch, and mesh, both models received the same rotation of the raw coordinates and inlet direction before preprocessing. We disabled their native augmentation routines because those sample different distributions. All runs used the locked 413/45/50 split, 100 epochs, and checkpoint selection by canonical validation error. Model-specific learning rates and node subsampling were retained, so compute was not matched. Training Model Canonical error Non-grid error Non-grid defect Canonical only CEGINO 0.3409 ± 0.0320 0.3402 ± 0.0319 0.0325 ± 0.0043 Canonical only EqGINO 0.1892 ± 0.0014 0.8843 ± 0.0037 0.8823 ± 0.0056 Shared augmentation CEGINO 0.3457 ± 0.0051 0.3448 ± 0.0047 0.0387 ± 0.0088 Shared augmentation EqGINO 0.2575 ± 0.0089 0.2264 ± 0.0091 0.2057 ± 0.0057

Values are means ± sample SD across three training seeds. Non-grid metrics average seven rotations on the same 50 meshes within each seed. These are fresh fixed-budget runs, separate from the original checkpoints.

The reviewer’s hypothesis is supported: augmentation reduces EqGINO’s non-grid error from 0.8843 to 0.2264 and reverses the accuracy ranking. The distinction is that CEGINO already transfers to the tested non-grid frames without seeing rotations during training. Even when both models receive augmentation, CEGINO retains an approximately 81% smaller non-grid equivariance defect (0.0387 vs. 0.2057). Its advantage in that setting is consistency across frames, not lower prediction error.

W3 — What do invariant descriptors preserve, and what might they discard? We tested whether restoring access to absolute orientation improves canonical accuracy. We appended either three zero channels or three centered Cartesian coordinates, matching input width, parameter count (1,204,361), and initialization within each seed. Coordinates were recomputed after rotation. Adding Cartesian coordinates did not yield a demonstrated canonical benefit: mean error changed from 0.3223 to 0.3392 (paired difference, 95% interval: −0.0387 to 0.0724). Yet non-grid error rose from 0.3215 to 0.7005, and defect from 0.0263 to 0.5862; both worsened in every seed. This gives direct evidence for the value of CEGINO’s descriptor choice: in this matched intervention, withholding absolute orientation protected cross-frame prediction accuracy. It does not show that all directional information is dispensable, or rule out improvements from richer relative geometry or chirality-sensitive features.

W4 — Figure 1 and the source of the advantage. The reviewer is correct: “Fourier Transform,” “Spectral Grid,” and “Orbit-based Weights” in Figure 1 do not describe the evaluated implementation. CEGINO lifts invariant surface features to sparse anchors, updates them by fixed-step local residual message passing, and decodes at query points using two spatial scales and pooled anchor context. The matched pressure configuration also uses one four-head attention block for communication across anchors. This distinction matters for the comparison with EqGINO: CEGINO can exchange information across the surface without mapping features onto a Cartesian Fourier grid. Its measured non-grid robustness is therefore a property of the tested anchor-based model, not evidence for a spectral component or a claim of exact numerical equivariance.

Q3 — Can a consistent prediction still be inaccurate? Yes. A model predicting the same incorrect constant pressure in every frame would have zero defect. For pressure predictions p0,pRp_0,p_R at corresponding nodes and unchanged target yy, the reverse triangle inequality gives

∣ER−E0∣≤DR ∥p0∥∥y∥,|E_R-E_0|\leq D_R\,\frac{\|p_0\|}{\|y\|},

where E0,ERE_0,E_R are relative prediction errors and DR=∥pR−p0∥/∥p0∥D_R=\|p_R-p_0\|/\|p_0\|, with nonzero denominators. Defect thus bounds the change in error, not its absolute level. This is why the canonical-only result is particularly informative: CEGINO had both smaller defect and lower error against the physical target on unseen non-grid rotations. After augmentation, only the consistency advantage remained.

Q4 — How often is the WSS basis nearly degenerate? CEGINO’s local basis lets invariant coefficients reconstruct a vector that rotates with the geometry, provided the basis is non-degenerate. We checked this condition at all 3,598,547 vertices of the 50 test meshes. The sine of the angle between the inlet and centered radial directions fell below 10−6,10−4,10−3,10−210^{-6},10^{-4},10^{-3},10^{-2} at 0, 0, 14, and 1,013 vertices, respectively. Near-alignment below 10−210^{-2} occurred on every mesh but affected only 0.02815% of vertices. For one archived WSS checkpoint, those vertices accounted for an average of 0.01949% of the squared equivariance residual energy per non-grid mesh–rotation pair (maximum 0.2914% across 350 pairs). Thus, near-aligned vertices were not the main source of the observed vector inconsistency in this diagnostic. The checkpoint’s original training configuration is incomplete, so this is a single-checkpoint result, not a guarantee for other geometries.

Q5 — Is the cross-frame benefit driven by a few favorable meshes or angles? In the original canonical-trained comparison, CEGINO had lower non-grid pressure error than EqGINO in all 1,050 paired seed–mesh–rotation observations (3 seeds × 50 meshes × 7 rotations). Thus the aggregate zero-shot advantage did not arise from a few unusually difficult EqGINO cases. We also generated mesh-by-angle heatmaps and curves for all 50 meshes in each seed under both fresh training conditions. The augmented comparison provides a useful counterpoint: for seed 0, EqGINO had lower error on every mesh after averaging its seven non-grid rotations. These counts describe the observed meshes; they are not independent training replicates. Together, they support a specific deployment conclusion: when training is restricted to a canonical frame but inference must handle unseen orientations, CEGINO offers more reliable non-grid pressure predictions with a smaller measured resource footprint.

Thank you for the detailed responses and additional experiments. The shared rotation-augmentation comparison addresses my main concern and clarifies the scope of CEGINO’s advantage. However, the provided manuscript still states that this comparison has not been conducted, and Figure 1 retains labels inconsistent with the implementation. Please incorporate the new results and mesh-by-angle visualizations into the paper and revise the figure accordingly. The canonical accuracy gap and the potential loss of azimuthal or chiral information remain concerns.

Thank you for the follow-up. We are glad the shared-augmentation comparison addressed your main concern. Our previous reply did not clearly distinguish the revisions we described from the PDF available to you. In the next permitted revision, we will replace the statement that this comparison has not been conducted with the three-seed results, report prediction error and equivariance defect side by side, and include mesh-by-angle heatmaps so readers can see the variation across all 50 test meshes.

The distinction we will make explicit is that CEGINO’s strength is consistency across coordinate frames, not uniformly lower prediction error. On AhmedBody pressure, without rotation augmentation, CEGINO’s mean relative L2 error was nearly unchanged from the canonical frame to non-grid rotations (0.3409 to 0.3402), whereas EqGINO’s increased from 0.1892 to 0.8843. With identical Haar rotations during training, EqGINO achieved lower non-grid prediction error (0.2264 vs. 0.3448), while CEGINO retained a lower equivariance defect (0.0387 vs. 0.2057). We will preserve both findings, including EqGINO’s canonical accuracy advantage.

Figure 1 will show the evaluated forward pass: invariant surface features are pooled onto sparse anchors, updated by local residual message passing, optionally mixed by global anchor attention, and decoded at query points using two neighborhood scales. We will remove the Fourier, spectral-grid, and orbit-weight labels, and distinguish scalar-field prediction from vector reconstruction in a rotating local basis.

Finally, the scalar descriptors can map distinct azimuthal or mirror-related configurations to the same feature values. Our Cartesian-coordinate ablation tested access to absolute pose; it did not isolate azimuthal information or chirality. We will state that limitation directly and will not attribute the canonical accuracy gap to a cause the experiments did not establish.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.