acceptodds
Under review as a conference paper at ICLR 2027

Mesh-Native Tokenisation: Retrofitting Foundation Models to Unstructured Data

Abstract

Physics foundation models are an increasingly popular approach for building versatile emulators for spatiotemporal systems. They require tokenisation layers that lift physical field data into a latent representation suitable for dynamic processing. A typical choice for this layer is a convolutional encoder, which lifts field information in a locality-aware way and can be trained very efficiently. However, this choice cannot be applied to a large class of scientific computing problems where data are represented on unstructured meshes, for example simulations on domains with curved boundaries. Such data typically require graph neural network architectures, which are more expensive to train and, in the form currently used, encode information differently from their CNN counterparts, thus limiting the portability of physics emulators between these data regimes. In this work we introduce Mesh-Native Convolution (MNC), a locality-aware tokenisation layer that encodes unstructured mesh-based data in a way that is analogous to CNNs. This approach leads to models that can be trained once on mesh data and then deployed to arbitrary unstructured meshes, even if these are irregular, with minimal retraining. We show that this tokenisation approach can be used to retrofit grid-pretrained foundation models efficiently on unstructured meshes and provide theoretical results to confirm the consistency of this approach with classical CNN encoders. We demonstrate the favourable performance of this retrofitting approach, which in many cases reaches acceptable accuracy even in zero-shot deployment, and outperforms existing baselines after only a few epochs of fine-tuning, thus providing an efficient bridge between the two data regimes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, Physics foundation models, Unstructured meshes, Transfer learning, Continuous convolutions

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kmKbOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Mesh-Native Convolutions (MNC), a tokenizer that enables models trained on grid data to process arbitrary meshes. After describing the MNC layers and briefly comparing them with GCN, they apply the method to Overtone. The model(s), originally trained on grid data, outperform mesh-native models such as GINO or LNO, even with zero-shot transfer. Similarly, combining MNC with Walrus, they model the cylinder flow dataset, a system that requires irregular meshes and that Walrus has never seen.

Critical strengths

Their work enables models trained on grid data to process arbitrary meshes without (or with very little) further training.

Critical weaknesses

I doubt experiments I and II are fair or useful. Except for LNO, the benchmark models are significantly smaller than the models MNC is built on. Even if size were similar, it's unclear whether MNC offers any benefit (maybe the model is just better), and why should it? If MNC enables models trained on grid data to process arbitrary meshes without a performance drop, I'd call it a success. If you show this (maybe by combining MNC with small models in prediction tasks), I'd be happy to increase my rating.

Minor criticism: given how MNC is defined, the exact consistency with CNN encoders seems more a valuable sanity check than a separate contribution.

Questions for authors
  • Related to the weakness, have you tried coupling overtone or other models to point-cloud encoders?

  • Out of curiosity, as I don't expect you to do it within the rebuttal time, have you tried your approach on 3D data? The extension seems pretty natural and would strengthen the value of your method.

3 comments

We thank the reviewer for their detailed reading of our work and we particularly appreciate the reviewer's concrete criterion for reconsidering their assessment: whether MNC can transfer grid-trained models to irregular meshes without a substantial loss of accuracy, including for smaller models. We have now tested precisely this, and added further comparisons and clarifications that directly respond to the concerns and questions raised by the reviewer.

The overall conclusion is that MNC moves grid-trained models onto meshes with very little loss of accuracy at every model size from 2M to 26M parameters, is more accurate zero-shot than a distilled point-cloud encoder, and extends to 3D. We believe this addresses the reviewer's concerns, and would thus warmly welcome an increase in their score. If the reviewer has remaining concerns we would invite them to raise those points explicitly for discussion. Please find a point-by-point reply below. Changes are also highlighted in red in the revised pdf.

Response to critical weaknesses:

(a) Mesh transfer without performance drop: This property was already tested and demonstrated in experiment 2, but we have included a more extensive evaluation in a new experiment in Appendix F to confirm the favourable transfer property. Here we train Overtone at three sizes on the TRL2D grid and evaluate it zero-shot on five meshes through MNC. Over ten-step rollouts the error stays within 7% of the grid on most meshes, at every size; and even on the highly graded mesh the error remains far lower than random weights. VRMSE (×10−210^{-2}):

ParametersGridFrontalIrregularJitteredFine 2×Graded
1.7M56.656.5 (+0%)56.6 (+0%)60.6 (+7%)57.7 (+2%)73.0 (+29%)
6.6M53.153.3 (+0%)54.2 (+2%)55.4 (+4%)53.9 (+1%)70.5 (+33%)
26.3M49.549.6 (+0%)52.3 (+6%)52.3 (+6%)51.5 (+4%)60.9 (+23%)
Persistence86.286.186.186.186.186.1
Random weights318.3336.8337.1333.5337.7335.0

The same holds with an FNO processor behind the tokeniser, so the transfer is not a property of one model but rather the encoder architecture:

ParametersGridFrontalIrregularJitteredFine 2×Graded
1.8M60.760.1 (−1%)62.2 (+3%)62.2 (+2%)60.8 (+0%)71.0 (+17%)
7.3M54.053.0 (−2%)55.0 (+2%)57.6 (+6%)56.1 (+4%)63.8 (+18%)
28.9M52.951.9 (−2%)56.5 (+7%)55.8 (+6%)53.1 (+1%)63.1 (+19%)
Persistence86.286.186.186.186.186.1
Random weights308.1323.7323.4320.5324.4320.6

(b) Exact consistency as a contribution: We agree, and have recast it as a design feature rather than a separate contribution (Sec. 1 and before Proposition 1). However, we politely disagree that it is only a sanity check: MNC was built around this property, and a mesh tokeniser need not have it. Point-cloud encoders, for example, have no grid counterpart to load weights from, so their tokeniser must be trained or distilled against the processor. With MNC the pretrained weights are loaded directly, and the saving is measurable: in Figure 6 the retrofit reaches its plateau within ten epochs of fine-tuning, against about a hundred from scratch.

Answers to questions:

  1. Point-cloud encoders: Yes. We coupled Overtone to a PointNet++ tokeniser (the better of two variants), which has to be distilled from the CNN tokeniser because it cannot load its weights. MNC is more accurate zero-shot on all three datasets, and needs no distillation. After fine-tuning, the distilled PointNet++ draws level on TRL2D and shear flow, and MNC stays ahead on active matter. Next-step / 10-step VRMSE (×10−210^{-2}) on the irregular meshes:
TokeniserTRL2DActive matterShear flow
MNC, zero-shot26.72 / 53.14.93 / 26.11.44 / 8.50
MNC, fine-tuned22.10 / 49.03.30 / 22.80.43 / 2.65
PointNet++ (distilled), zero-shot33.44 / 79.611.73 / 53.91.61 / 8.58
PointNet++ (distilled), fine-tuned22.40 / 49.54.17 / 25.20.46 / 2.46
PointNet++ (random init), fine-tuned24.54 / 58.96.57 / 31.00.74 / 4.15

A targeted comparison with the same Walrus checkpoint is the new Experiment 4 (Sec. 4.4).

  1. 3D data: Yes, we have now implemented it: the weights of a Conv3d transfer in the same way. We are evaluating the architecture on the 3D MHD dataset of the Well; the runs (close to 10,000 USD of compute so far) are still ongoing. Interim results, from a 64364^3 grid to a 261k-node tetrahedral mesh, show the same picture as in 2D. Next-step / 10-step VRMSE (×10−210^{-2}):
ModelZero-shot on the meshAfter training on the mesh
Overtone+MNC, retrofit from the grid21.52 / 62.6811.69 / 50.36
Overtone+MNC, from scratch-13.02 / 51.15
Transolver++31.12 / 84.0529.38 / 98.45
LNO38.52 / 92.7136.98 / 92.71

These come from early checkpoints of the grid pretraining for a single seed; we will add the final results to the revised manuscript as they arrive. We thank the reviewer for this excellent suggestion.

Thank you for the detailed reply and additional work. I appreciate the new experiments on different meshes and the preview of the 3D results. I would also like to thank Reviewer S8Rj for asking about direct alternatives, and the authors for carrying out that comparison. It is a valuable addition.

As my concerns have been addressed, I raise my score to 3.

Two suggestions that wouldn't change the rating:

  • The experiments in Appendix F deserve a place in the main text. I understand that space is limited, but at least Table 11 would be worth adding.
  • With alternatives in the main paper, it would make sense to mention them in the discussion. The strong performance of interpolation is worth noting.

We thank the reviewer for their positive response and for taking the time to carefully consider our rebuttal and additional experiments. We are very pleased that these have addressed their concerns, and greatly appreciate their constructive feedback.

Score update: We are very grateful for the reviewer's decision to increase their rating to 3 (weak acceptance). We noticed, however, that the official review still displays the original rating of 1 (clear rejection), so we wondered whether the change might not have been saved. Would the reviewer mind checking this once more? We apologise for the additional request and very much appreciate their help (scores can be updated by editing the official review on OpenReview).

Suggestions: We also thank the reviewer for their two additional suggestions, both of which we have incorporated into the revised manuscript:

  1. Mesh transfer experiments in the main text: We have now included a concise presentation of the mesh transfer experiments at the end of Sec. 4.1, with a discussion pointing to the more extensive results in Appendix F, including the FNO experiments.

  2. Discussion of alternative tokeniser performance: We have expanded the discussion of Experiment 4 (Sec. 4.4) to acknowledge the strong performance of interpolation and highlight the regimes in which MNC outperforms interpolation and the other tested alternatives.

Finally, our 3D experiments are now included in Appendix E.5, together with representative figures, demonstrating that MNC also retains its strong performance in the 3D setting.

Once again we thank the reviewer for their time and consideration.

Reviewer_S8RjOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Mesh-Native Convolution (MNC), a mesh-compatible tokenization framework that transfers convolutional encoder–decoder weights from grid-pretrained physics models to unstructured meshes while retaining their pretrained processors. The method combines continuous spline kernels, occupancy normalization, and positional-interface adaptation, supported by conditional grid-equivalence proofs and forecasting experiments on four physical datasets.

Critical strengths

Practical Reuse of Pretrained Models: MNC addresses the interface mismatch between grid-pretrained physics models and unstructured inputs. Direct weight transfer provides a practical alternative to retraining mesh-native models from scratch.

Positive Transfer Evidence: Zero-shot and fine-tuning experiments (Table 1,2,3) demonstrate useful transfer across three resampled physical datasets and the cylinder-flow benchmark. Scratch controls and processor-free reconstruction further support the value of the transferred representations.

Critical weaknesses

Insufficient Comparison with Direct Alternatives: The paper acknowledges prior continuous-kernel interpolation and native-grid recovery (Sec1.1), but does not experimentally establish its advantage over these approaches. A mesh-to-grid resampling baseline using the same pretrained checkpoint is also absent, leaving the necessity and incremental benefit of MNC insufficiently justified.

Limited Support for Broad Mesh Generalization: The theoretical limit retains dependence on the sampling density ρ\rho (Proposition 2), while the Well experiments use matched-resolution meshes without physics-based grading. The Walrus experiment additionally requires substantial mesh refinement and mass-weighted aggregation, limiting support for general claims of mesh-agnostic deployment.

Questions for authors

Technical Distinction from Prior Work: Given the prior filter interpolation and native-grid recovery acknowledged in Sec. 1.1, which specific design choices distinguish MNC, and which submitted results isolate their contribution?

Necessity of Mesh-Native Adaptation: How would MNC compare with interpolating mesh inputs onto a regular grid, running the same pretrained checkpoint, and interpolating predictions back? This comparison would clarify whether MNC provides benefits beyond straightforward checkpoint reuse.

3 comments

We thank the reviewer for their careful review and for the suggestions on clarifying the novelty and significance of this work, which led us to add experiments comparing MNC with direct alternatives and demonstrating its generalisation across meshes.

The overall conclusion is that MNC outperforms both interpolation with the same checkpoint and prior continuous-kernel work, and transfers across meshes of different resolution and grading. We believe this addresses all concerns raised by the reviewer, and would thus warmly welcome an increase in their score. If the reviewer has remaining concerns we would invite them to raise those points explicitly for discussion.

Please find a point-by-point reply below. Changes are also highlighted in red in the revised pdf.

Response to critical weaknesses:

(a) Comparison with Direct Alternatives: We have added this comparison as a new Experiment 4 (Sec. 4.4). The pretrained Walrus checkpoint is kept fixed and only the tokeniser is changed: MNC; interpolation of the mesh onto the grid and back, exactly as the reviewer suggests; the node weighting of DI-Net (Wang & Golland, 2023); and a PointNet++ point-cloud tokeniser (Qi et al., 2017). MNC is the most accurate in every rollout metric, and its advantage over interpolation is largest near the cylinder wall. Representative results are shown below (VRMSE ×10−210^{-2} on cylinder flow); full results are in the revised manuscript.

Walrus withNext step10 steps50 stepsNear wall, steps 11-50Wake, steps 11-50
MNC0.15 ± 0.011.02 ± 0.004.77 ± 0.694.80 ± 0.068.61 ± 0.38
Interpolation0.15 ± 0.001.06 ± 0.045.45 ± 0.366.54 ± 0.729.72 ± 0.90
DI-Net weighting0.19 ± 0.001.42 ± 0.128.15 ± 0.377.57 ± 0.2318.48 ± 1.24
PointNet++0.24 ± 0.111.71 ± 0.0211.64 ± 0.3211.32 ± 0.3425.85 ± 0.78

(b) Mesh Generalisation: We ran two new sets of experiments and address each point below.

  • Other resolutions and gradings (Appendix F): grid-trained models, moved zero-shot onto five meshes (including one with twice the nodes and one graded in density) for two processors and three sizes, stay close to the grid error and demonstrate significant transfer ability. 10-step VRMSE (×10−210^{-2}) for the largest models:
GridFrontalIrregularJitteredFine 2×Graded
Overtone+MNC49.549.6 (+0%)52.3 (+6%)52.3 (+6%)51.5 (+4%)60.9 (+23%)
FNO+MNC52.951.9 (−2%)56.5 (+7%)55.8 (+6%)53.1 (+1%)63.1 (+19%)
Persistence86.286.186.186.186.186.1
  • No refinement needed (Appendix E.3): our new high-Re meshes have 0.38–0.84M nodes, graded to 0.3–0.6 mm at the wall. Walrus+MNC runs on them directly, without refinement or retraining, at 40% of the error of persistence. Experiment 3 refined the original meshes only because their ~1.9k nodes are coarser than Walrus's token windows.
  • Density dependence: we politely disagree with the characterisation of the node-density dependence of MNC's continuum limit as a restriction. Indeed this is a natural property of any layer that aggregates over mesh nodes: where a mesh places more nodes, the field there carries more weight. Proposition 2 shows that this is the only way the mesh enters, so the output does not depend on connectivity, unlike GCN-type layers (Fig. 4b).

Answers to questions:

  1. Technical Distinction from Prior Work: We appreciate the feedback and have clarified this in Sec. 1 and 1.1. SplineCNN learns its kernels from scratch; only DI-Net transfers pretrained filters, weighting all sample points equally. MNC moves a complete pretrained model onto an irregular mesh, through three design choices and contributions:
  • Consistent representation: Proposition 2 shows that mesh connectivity and individual node positions do not matter in the limit, only the sampling density (Fig. 4), which is exactly what a tokeniser for physical fields should offer. This is demonstrated in Experiment 4, and Appendix A.4.
  • Weight transferability: a coordinate-based rotary embedding lets a grid-pretrained processor run unchanged on mesh tokens (Appendix A.2) which leads to the excellent the zero-shot results of Experiments 1 and 3.
  • Encoder-decoder hierarchy: strided and transposed convolutions are reproduced exactly at every level (Proposition 1), and the tokeniser shows excellent grid to mesh transfer properties making it a physically valid encoder architecture (cf. Appendix F).
  1. Necessity of Mesh-Native Adaptation: The key insight is that MNC combines the flexibility of consistent convolutional representation, with the resolution accuracy of graph neural networks and thus outperforms other mesh-based encoder architectures, as is particularly visible in Experiment 4 and Figure 20. Summary of results:
Next step50 stepsNear wall, steps 11-50
MNC0.15 ± 0.014.77 ± 0.694.80 ± 0.06
Interpolation0.15 ± 0.005.45 ± 0.366.54 ± 0.72
Reviewer_S8RjOpenReview ↗

Thanks for the new experiments

Thank you for the detailed response and the additional experiments. These additions have addressed most of my concerns, particularly regarding comparisons with direct alternatives and generalisation across mesh resolutions and gradings. I have therefore raised my score.

We thank the reviewer for their positive response and for taking the time to carefully consider our rebuttal and additional experiments. We are very pleased that these additions have addressed their concerns, and we greatly appreciate their constructive feedback throughout the review process.

Reviewer_exkeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents the Mesh-Native Convolution (MNC) method a locality-aware tokenisation layer that encodes unstructured mesh-based data in a way that is analogous to CNNs. It adapts pretrained physics models from regular grids to unstructured meshes. The idea is to reuse the weights of existing models instead of training a new model from scratch. The method is tested on several numerical tests, including cylinder flow showing prediction accuracy and faster fine-tuning.

Critical strengths

The method addresses an interesting question related to CFD on unstructured meshes: transferring pretrained models to unstructured meshes without retraining everything from scratch. The experiments cover several datasets and include both one-step predictions and autoregressive rollouts. The cylinder-flow test shows that the method can transfer a pretrained model to a new flow configuration.

Critical weaknesses

The evaluation uses only field-based RMSE. It does not analysis whether important physical quantities, such as drag, lift, or vortex-shedding frequency, are also predicted accurately. The rollout evaluation is limited to ten steps, so the long-term stability of the predictions remains unclear. Some comparisons are difficult to interpret because the proposed method uses pretrained weights, while several baselines are trained from scratch.

Questions for authors

Can the method handle high-Reynolds-number flows with thin boundary layers, i.e. cylinder flow at high Reynolds number? Can the authors report errors in drag and lift coefficients, as well as their time histories, to assess whether the predictions preserve the relevant flow physics? How well does the method perform over longer autoregressive rollouts, beyond ten steps? Can the authors evaluate errors near the cylinder wall and in the wake separately, rather than relying only on a global error metric?

2 comments

We thank the reviewer for their careful consideration of our submission and for their comments. Based on their suggestions we have conducted extensive further numerical experiments (equivalent to approx. USD 10,000) to provide for a more comprehensive evaluation of our novel method.

The overall conclusion is that in nearly all additional metrics and datasets proposed by the reviewer our method significantly outperforms the baselines. We believe this answers all the questions and weaknesses raised by the reviewer, and would thus warmly welcome an increase in their score. If the reviewer has additional concerns related to their score we would invite them to explicitly raise those points for discussion. Please find a point-by-point reply below. Changes are also highlighted in red in the revised pdf.

Answers to questions:

  1. High-Reynolds-number flows: Yes, our method performs very well also on high-Reynolds-number flows. The original evaluation dataset is the publicly available Cylinder Flow dataset from MeshGraphNets, which contains trajectories of Re∈[14,212]\mathrm{Re}\in[14,212], so our previous evaluation already included cases of moderate Reynolds number. Nevertheless, we have generated a full new high Reynolds number dataset with Re\mathrm{Re} ranging from 567567 to 19821982. Our MNC encoder coupled with Walrus outperformed all baselines on this new dataset; full results are in Appendix E.3, with a representative table below.
Metric (steps 1-50)Walrus+MNC, zero-shotBest baseline
VRMSE (×10−210^{-2})40.53102.08 (persistence)
CLC_\mathrm{L} error (%)54.5114.4 (persistence)
CDC_\mathrm{D} error (%)15.9415.86 (persistence)
St\mathrm{St} error (%)14.839.7 ± 4.2 (Transolver++)
Wall pressure error (%)48.173.1 ± 0.8 (LNO)
Near-wall VRMSE (×10−210^{-2})49.7492.62 (persistence)
Wake VRMSE (×10−210^{-2})48.42138.52 (persistence)
Far-field VRMSE (×10−210^{-2})32.3774.60 (persistence)
  1. Errors in drag and lift coefficients: We use VRMSE as the de-facto standard pointwise metric for PDE surrogates, but lift (CLC_\mathrm{L}), drag (CDC_\mathrm{D}) and the Strouhal number (St\mathrm{St}) can also be computed. We have included those results together with histories in Appendix E.2 and note that our encoder combined with Walrus outperforms all baselines in lift and Strouhal number (drag is nearly constant in time, so every model is within 3%). Relative errors over steps 1-50:
ModelCLC_\mathrm{L} (%)CDC_\mathrm{D} (%)St\mathrm{St} (%)
Walrus+MNC, pre-trained, fine-tuned36.3 ± 6.30.78 ± 0.116.2 ± 0.7
Walrus+MNC, pre-trained, zero-shot97.10.9857.6
MeshGraphNets165.9 ± 3.52.21 ± 2.2088.7 ± 23.8
Transolver++122.0 ± 17.72.61 ± 0.6136.8 ± 7.3
GINO65.8 ± 2.50.55 ± 0.1216.2 ± 12.2
LNO114.8 ± 0.00.48 ± 0.01–
Persistence114.80.47–
  1. Longer autoregressive rollouts: Thank you, we have now re-evaluated all of our numerical experiments on horizons extending to 50 steps and included those results in Tables 1-3 of the manuscript (and in Appendix E). Throughout our method outperforms baselines also over these horizons, as demonstrated on the following representative Table from experiment 3.
Model50-step VRMSE
Walrus+MNC, pre-trained, fine-tuned4.77 ± 0.69
Walrus+MNC, pre-trained, zero-shot20.38
Walrus+MNC, random init, fine-tuned9.75 ± 0.67
MeshGraphNets, trained on cylinder flow842 ± 1140
Transolver++, trained on cylinder flow16.68 ± 0.25
GINO, trained on cylinder flow12.75 ± 2.68
LNO, trained on cylinder flow21.95 ± 0.33
Persistence21.62
  1. Errors near cylinder wall and wake: This is very much possible, and we have now included additional metrics for evaluation (together with sketch) in Appendix E.4 of the revised manuscript. Results over steps 1-50 (×10−210^{-2}):
ModelWall VRMSEWake VRMSE
Walrus+MNC, pre-trained, fine-tuned3.79 ± 0.327.36 ± 0.34
Walrus+MNC, pre-trained, zero-shot13.7433.26
MeshGraphNets31.68 ± 6.8343.50 ± 5.35
Transolver++23.35 ± 5.9137.71 ± 13.14
GINO8.49 ± 1.2626.04 ± 8.47
LNO10.86 ± 0.0046.85 ± 0.00
Persistence10.8646.85

Response to critical weaknesses:

(a) Evaluation metrics and rollout horizons: Please see the above answers.

(b) Interpretation of results pretrained vs trained from scratch: We would like to clarify that the only comparison in the paper of pretrained models with re-trained baselines is in Experiment 3, and would like to note that this is indeed a particular strength of our methodology: the entire point being that we can reuse a grid-pretrained model on unstructured meshes. Table 3 also reports the same architecture trained from random initialisation, which isolates the effect of pretraining.

Dear Reviewer exke,

Thank you again for your careful review and constructive feedback. As the discussion period draws to a close, we would appreciate your feedback on whether our detailed rebuttal and revised manuscript have addressed your concerns.

In particular, we have added extensive new experiments on high-Reynolds-number flows, 50-step autoregressive rollouts, physical quantities (lift, drag, and vortex shedding), and errors near the cylinder wall and in the wake. These new results demonstrate that our method performs favourably against the baselines across all additional metrics and settings you suggested.

We believe these additions directly address the concerns raised in your review and would greatly appreciate your consideration of an updated score should you feel the new results warrant it. If any concerns remain unresolved, we would be grateful if those could be raised explicitly so that we can provide further clarification before the end of the discussion period.

Thank you again for your time and consideration!

Best wishes,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to AC and all reviewers

Dear Area Chair and Reviewers,

Thank you for your careful reviews and for the concrete and constructive feedback. We have revised the manuscript to address the concerns raised, and the updated version has been uploaded to OpenReview with all changes highlighted in red. The main updates are summarised below.

  • Comparison with alternative tokenisers. We added a new Experiment 4 (Sec. 4.4, Appendix G), in which the same pretrained Walrus checkpoint is combined with four tokenisers: MNC, interpolation onto the grid and back, DI-Net, and a distilled PointNet++ point-cloud tokeniser. Errors are reported over the whole field, near the cylinder wall and in the wake.
  • Transfer across meshes and model sizes. We added Appendix F, in which two models (Overtone, and an FNO processor behind the same tokeniser), each at three sizes from 2M to 29M parameters, are trained on the grid and evaluated zero-shot through MNC on five meshes of different connectivity, resolution and grading.
  • Physical quantities and local errors. We added lift, drag and Strouhal-number errors with their time histories on cylinder flow (Appendix E.2), and errors resolved near the cylinder wall, in the wake and in the far field (Appendix E.4).
  • Longer rollouts. We extended all experiments to 50-step autoregressive rollouts (Tables 1-3).
  • High-Reynolds-number flows. We generated a new cylinder-flow dataset with Reynolds numbers between 567 and 1982 and evaluated every model on it without retraining (Appendix E.3).
  • Contributions and related work. We clarified the contributions and the distinction from prior continuous-kernel work (Sec. 1 and 1.1), presenting exact consistency with CNN encoders as a design feature that enables direct weight transfer, and added a description of the positional interface for mesh tokens (Appendix A.2).

Detailed responses to the individual concerns are provided in the corresponding reviewer threads. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.