acceptodds
Under review as a conference paper at ICLR 2027

StarVortex: Self-Supervised Geometric Representation Learning of Collective Integral Curves for Lagrangian Vortices

Abstract

Learning useful representations of dynamical systems from geometry alone is challenging: individual trajectories provide limited information, while the relevant structure often lies in the collective deformation of neighboring trajectories and should be insensitive to arbitrary spatial orientation. We study this problem in fluid flows and introduce StarVortex, a self-supervised framework for learning geometric representations directly from local bundles of integral curves. We represent each neighborhood as a star consisting of a center curve and symmetrically seeded neighbors, exposing both center-curve transport and relative bundle deformation. Rather than requiring a fully rotation-equivariant architecture, StarVortex combines an exact architectural bias induced by the discrete symmetry of the sampling stencil with a representation-level objective that enforces consistency under arbitrary continuous rotations. Reconstruction and variance–covariance regularization further encourage the latent space to retain informative geometric variation. The resulting representation captures vortex-relevant structure without velocity, vorticity, physics-derived descriptors, or training labels. It supports label-free discovery of vortex regions and provides transferable features for supervised recognition of vortex regions, corelines, and hairpin vortices across pathlines, streamlines, and vortex lines. Experiments and ablations show that explicitly structuring self-supervision around geometric symmetry is critical for learning useful representations, and that the learned features are particularly beneficial when labels are scarce or evaluation shifts beyond the training domain.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, vortex

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_FaPiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents StarVortex, a framework for learning geometric representations from local bundles of integral curves. The method constructs star-shaped neighborhoods of curves and combines shared neighbor processing with rotation-consistency learning, reconstruction, and representation regularization. The learned features are evaluated on unsupervised vortex-region clustering and supervised recognition of vortex regions, corelines, and hairpin vortices.

Critical strengths
  1. Meaningful geometric representation. Learning from neighboring integral curves is well motivated, particularly when trajectory data are available, but full velocity fields are expensive or unavailable. The unsupervised vortex-region results show substantial improvements over the provided geometric clustering baselines.

  2. Promising empirical performance. The coreline experiments demonstrate strong recognition accuracy, including on genuinely unseen flows. The limited-label and domain-generalization studies also provide evidence that the auxiliary self-supervised objectives can improve supervised learning.

Critical weaknesses
  1. Limited methodological novelty and unclear attribution of improvements (see Q1). The proposed hard symmetry mainly relies on conventional shared neighbor processing, which provides permutation equivariance, while the soft symmetry uses rotation-consistency learning. The paper does not sufficiently establish the group-specific contribution of its architecture or demonstrate whether the improvements arise from the proposed symmetry design rather than conventional geometric encoding and augmentation.

  2. Insufficient validation of the latent decomposition (see Q2). Rotation consistency is imposed only on (zinvz_{\mathrm{inv}}), while reconstruction uses the full latent code. Although this encourages an invariant representation, it does not guarantee that the two components are meaningfully disentangled. The network could preferentially retain useful geometric information in (zposez_{\mathrm{pose}}), and the current experiments do not directly examine this possibility.

  3. Test-based model selection (see Q3). Appendix C.3 explicitly selects coreline checkpoints using test macro-F1, compromising the independence of the original test evaluation. Although the three additional unseen flows are excluded from checkpoint selection, the aggregated results may still include selection-affected test data.

  4. Limited evidence for broader label-free applicability. Unsupervised vortex-region identification trains and clusters each flow independently, leaving cross-flow transfer of the learned representation untested in this setting. Figure 5 seems to demonstrate geometric-feature clustering rather than the learned StarVortex model, so it does not directly validate the proposed representation on irregular particle trajectories.

Questions for authors

Q1. Symmetry design and improvement attribution.

What group-specific architectural property is introduced beyond conventional permutation-equivariant parameter sharing? Could the authors provide controlled comparisons with standard shared-neighbor encoding and rotation augmentation/consistency, and better position the method relative to recent rotation-equivariant approaches, such as ReViT (ICML 2026)?

Q2. Latent decomposition.

How do the authors ensure that useful geometric information is retained in (zinvz_{\mathrm{inv}}) rather than primarily encoded in (zposez_{\mathrm{pose}})? Could the two components and their concatenation be evaluated separately through downstream probes?

Q3. Coreline evaluation protocol.

Could the authors repeat the coreline comparison using validation-based checkpoint selection or a predefined training schedule? Please report the original and fully unseen flows separately under this protocol.

1 comment

Thank you for the constructive suggestions. We have added controlled comparisons and revised the evaluation protocol.

Symmetry design and attribution. We agree that permutation-equivariant parameter sharing alone is not specific to the stencil group. Table 7 now separates shared encoding, rotation augmentation, rotation consistency, arbitrary permutations, and exact stencil-group permutations. It also compares alternative stencils and neighborhood sizes, with configurations detailed in Appendix C.5. We additionally introduce REViT baselines in Tables 3, 5, 8, and 9, and describe the 3D adaptation in Appendix C.3.

Latent decomposition. Table 7(A) freezes the same encoder and applies identical linear probes to zinv, zpose, and their concatenation. The invariant code substantially outperforms the pose code, showing that useful task information is retained in zinv. These results support its use as the task descriptor without implying complete disentanglement.

Coreline evaluation protocol. We newly annotated Trefoil knot and combined it with steady tornado as an independent validation set. Checkpoint selection now uses validation macro-F1, with test sets excluded from training and model selection. Table 5 reports aggregate performance across the eight test flows, excluding validation flows. Table 9 provides per-flow results and identifies the three fully unseen flows, while Table 13 specifies the revised split.

Label-free applicability and particle data. The updated Fig. 5 directly evaluates learned StarVortex representations with GMM clustering on irregular particle trajectories, alongside vorticity visualizations used only for comparison. Our self-supervised setting learns and clusters each flow without labels; it does not establish transfer of a frozen encoder without retraining. The supervised coreline experiments separately evaluate transfer to fully unseen flows.

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

This paper investigates how to learn vortex-related representations from the geometry of local bundles of integral curves. The authors propose StarVortex, which constructs a star-shaped neighborhood from a center curve and symmetrically seeded neighboring curves, describing local flow through center-curve motion and relative deformation of the neighbors. The model uses shared neighbor processing, convolutional layers, and a Transformer encoder, and is trained with rotation consistency, reconstruction, and representation regularization. The resulting representations support unsupervised vortex-region clustering and supervised identification of vortex regions, corelines, and hairpin vortices.

Critical strengths
  1. The problem is practically relevant, and the curve-bundle representation is well motivated. Compared with individual trajectories, the relative deformation of neighboring curves provides richer information about local flow. Analyzing stored trajectories also suits applications in which storing complete time-dependent velocity fields is difficult.

  2. The distinction between the two types of symmetry is clear. The authors explicitly distinguish neighbor-index permutations from Cartesian vector rotations and acknowledge that the complete encoder is not strictly rotation equivariant. This is more accurate than interpreting parameter sharing as full rotation equivariance.

  3. The experiments cover multiple tasks and report several important limitations. Unsupervised 3D vortex-region identification achieves an F1 score of 0.6724, substantially outperforming the compared handcrafted geometric representations. Three additional flows excluded from checkpoint selection also provide encouraging evidence of coreline generalization. The authors additionally report that the unsupervised method trails the strongest supervised detectors and performs relatively poorly on Channel flow.

Critical weaknesses
  1. Appendix C.3, lines 1335–1340, explicitly states that the original protocol selects checkpoints every 200 updates based on pooled test macro-F1. This means that the original test set participates in model selection, and the aggregate results in Tables 5 and 8 that include these flows cannot be interpreted as an independent evaluation of generalization.

  2. Appendix C.4, lines 1348–1351, states that StarVortex uses a center curve and 12 neighboring curves in the hairpin experiments, whereas the convolutional and PointNet baselines retain a seven-curve representation. Thus, the comparison changes both the model and the available neighborhood information. This is particularly important because the paper’s motivation is precisely that multiple neighboring curves provide richer structural information. It is currently unclear how much of the improvement comes from the additional curves and how much comes from the symmetry design.

  3. The neighbor-permutation equivariance in Equation (6) follows from applying the same mapping to all neighbors. This property actually holds for arbitrary neighbor permutations and is not specific to D_8 or I_h. The authors therefore need to clarify the additional benefits provided by the finite stencil group and how these differ from ordinary shared neighbor encoding and rotation augmentation.

  4. A network receiving only curve geometry and an entire extraction pipeline requiring no physical-field information are two different claims. Hairpin identification uses vortex lines, whose generation requires a vorticity field; the surface reconstruction in Appendix C.4 additionally uses a local \lambda_2 threshold calibrated on training data. Therefore, the complete surface reconstruction results in Figure 3 do not, by themselves, establish that the entire pipeline relies only on particle positions. I recommend clearly distinguishing curve-bundle classification results from post-processing results that depend on physical fields, and specifying exactly what Volume IoU evaluates.

Questions for authors
  1. Could coreline checkpoint selection be repeated using an independent validation set, with results aggregated separately for the original flows and the three completely unseen flows?
  2. How much of the advantage in coreline and hairpin identification remains when all models use the same number of curves, sampled points, and normalization procedure?
  3. Which self-supervised losses, regularization terms, and representation dimensions are actually used for each task? With the variance-floor penalty disabled, how is degeneration of the invariant code assessed?
  4. Could you provide direct StarVortex results on the scattered particle data in Figure 5 and distinguish geometric classification from preprocessing or post-processing that depends on vorticity or \lambda_2?
2 comments

Thank you for identifying these evaluation and presentation issues. We have revised the protocols and clarified the distinction between geometric prediction and field-dependent processing.

Coreline evaluation protocol. We newly annotated Trefoil knot and combined its 15 frames with steady tornado to form an independent validation set. All coreline methods now select checkpoints using validation macro-F1, with the test set excluded from training and model selection. Table 5 reports aggregate performance across the eight test flows, excluding validation flows. Table 9 provides per-flow results and identifies the three fully unseen flows, while Table 13 specifies the revised split.

Matched curve counts. We reran the hairpin comparisons with 13 curves per bundle for all methods, replacing the previous seven-curve baselines. StarVortex retains the highest F1 across the three test flows in Table 10. Appendix C.4 describes the shared curve budget and each method’s input processing.

Contribution beyond parameter sharing. We agree that shared neighbor processing provides equivariance to arbitrary neighbor permutations. The finite stencil group supplies geometrically matched rotations and neighbor correspondences. Table 7 now compares ordinary shared encoding, rotation augmentation, rotation consistency, arbitrary permutations, and stencil-group permutations, together with alternative neighborhood templates. Appendix C.5 specifies these controls.

Geometry-only classification versus surface reconstruction. We now distinguish these stages explicitly. The classifier receives curve geometry, whereas hairpin vortex-line generation uses vorticity and surface reconstruction uses a locally calibrated 2 threshold. Appendix C.4 presents the reconstruction procedure in four steps, from filtering isolated predictions to removing overlapping surfaces.

Volume IoU. Thank you for pointing out this omission. The Volume IoU reported in Fig. 7 measures the intersection-over-union between the union of reconstructed hairpin volumes and the union of human-annotated reference volumes, computed on the native grid with physical cell-volume weighting. We have rewritten Appendix C.4 to detail surface reconstruction and IoU calculation, clarifying that this metric evaluates the complete reconstruction pipeline, including the 2-based post-processing.

Task-specific objectives and representations. Appendix C.1 and Table 11 summarize the readout dimensions and optimization settings, with task-specific objectives detailed in the subsequent subsections. Table 7 additionally compares downstream probes of the invariant code, pose code, and their concatenation.

Direct results on scattered particles. The updated Fig. 5 shows learned StarVortex representations clustered by GMM directly from irregular particle trajectories. Our earlier use of “geometric features and a GMM” may have caused confusion with traditional handcrafted geometric features; we apologize for the ambiguity and have revised the description. We also added vorticity visualizations for qualitative assessment. These are not inputs to the model or clustering, and the particle-data pipeline requires no velocity-field reconstruction.

Task-specific objectives and representation dimensions. We have clarified the distinction between the self-supervised vortex-region experiments and the supervised vortex, coreline and hairpin experiments in Table 2(b). By default the supervised experiments use only cross-entropy CE loss unless specified e.g. Table 6, although we show that the auxiliary unsupervised losses can help particularly with limited labels and under domain shift. For unsupervised tasks, the training objective combines orbit loss, reconstruction loss, and VICReg regularization. The encoder produces a 28-dim representation, partitioned into a 16-dim invariant code and a 12-dim pose code. The 16-dim invariant code is default source of readout unless specific notice. Hairpin classification is additionally tuned to uses the full 28-dim code, with weighted cross-entropy, multi-length Jensen–Shannon consistency, and reconstruction. We additionally clarify the regularization terms and their task-specific coefficients in the revised manuscript.

AuthorsOpenReview ↗

extending reply(due to characters limit)

Degeneration of the invariant code. The variance-floor penalty is enabled in our full-model experiments. To directly assess degeneration, we conducted an additional controlled ablation on Cylinder2D and Boussinesq, using identical initialization and measuring the invariant code without labels every 25 epochs. We report the results at 3,000 epochs below. Without the variance floor, the invariant code collapses rapidly: by epoch 25, spread falls to 0.088/0.091 and angular dispersion to 0.0051/0.0041 on the two datasets, respectively. At convergence, spread is reduced by 6–11× and angular dispersion by 32–68×. Notably, the orbit-consistency loss becomes lower despite representation collapse, demonstrating that invariance can be satisfied by a nearly constant code. Removing covariance regularization additionally has negligible effect, confirming that covariance alone cannot prevent degeneration. The resulting clustering becomes severely imbalanced (only 1–3% of stars in the smaller cluster), and macro-F1 decreases substantially. Thus, spread and angular dispersion directly detect invariant-code collapse without labels, while the variance floor prevents this trivial solution in the tested setting.

Table R1: Effect of variance-floor removal on invariant-code degeneration.

DatasetConfigurationSpread ↑Angular dispersion ↑Smaller clusterOrbit loss ↓Macro-F1 ↑
Cylinder2DFull0.5140.04712.9%1.8×10⁻⁴0.943
Floor off0.0470.00071.4%≈2×10⁻⁵0.567
Floor + cov. off0.0480.00071.4%≈2×10⁻⁵0.567
BoussinesqFull0.5250.06312.1%3.2×10⁻⁴0.829
Floor off0.0830.00202.7%1.2×10⁻⁴0.708
Floor + cov. off0.0860.00212.8%1.2×10⁻⁴0.718

Spread is the mean per-dimension standard deviation; angular dispersion is one minus the mean pairwise cosine similarity. Both are label-free. Macro-F1 uses labels only for evaluation. Results use a single seed of the 166k-parameter 2D model (8D invariant code).

Reviewer_JZ1POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a representation learning method for Lagrangian flows - whether integral curves computed on vector fields, or particles directly given. At a given point in the domain, an integral curve is computed, along with integral curves from seeds positioned at neighboring positions, arranged in an octagonal/icosahedral stencil. Latent representations are formed for all integral curves, with latent dimensions segmented into a rotation-invariant part, and the remaining bit. The center curve's rotation-invariant latent code is (mostly) used for downstream analysis, whether in an unsupervised way (a GMM fit to all latents), or in a supervised setting whereby the proposed self-supervision criteria supplements a cross-entropy loss for classifying a variety of vortex structures.

The main contribution of the paper is a self-supervisory objective for learning invariant latent representations of integral curves, aiming to satisfy two criteria: (1) a stencil-induced symmetry constraint, essentially matching integral curves that are related under elements of the symmetry group (permutation to select neighbor integral curve, followed by a rotation), and (2) rotation invariance.

The paper validates the approach by showing how the learned latent codes are predictive of different types of vortices: vortex regions, corelines, and hairpin vortices.

Critical strengths

The main strength of the paper is the novelty of the approach - the formulation proposed in learning representations over integral curve bundles, and the proposed "orbit" objective (c.f. Eq. 9) that encourages integral curves to match under actions of the symmetry group, and more general rotation invariance. Ablations in Table 4 convincingly show the benefits of both of these criteria. The informativeness of the invariant latent codes is further demonstrated quite well in the unsupervised setting (GMM fit, c.f. Table 3), as well as its ability to boost performance in limited-label settings (Table 6).

A secondary strength of the paper is its generality, namely the ability to support Lagrangian data. This is advantageous for particle data, whether collected experimentally or simulated via Lagrangian particle tracking. Although the evaluation is somewhat limited in this regard - qualitative results via Fig. 5 and accompanying video - the ability to extract, and track, vortex regions in particle data is noteworthy.

Critical weaknesses

A central weakness of the paper is that the experiments are rather contrived. Specifically, vortex regions and corelines are derived based on existing, fairly conventional flow feature methods, namely IVD [Haller et al. 2016] and parallel vectors [Reikert & Roth 1999], respectively. This poorly motivates the use of Lagrangian representation learning: why not just use these existing methods for vortex extraction to begin with? A distinction lies in parameter tuning and postprocessing to derive "cleaner" vortex structures. But to the best of my knowledge, the threshold used for IVD is not reported, and it is unclear how difficult it is to set in practice (and how different settings might impact the performance of the proposed method). Moreover, it appears from Table 5 that parallel vectors (what I take as the "raw" output, not post-processed), is already quite close to ground-truth, although it makes little sense that F1 and Chamfer distance are both so high. Unfortunately, Table 5 is never referenced in the paper, so I might be misinterpreting these results.

A related issue lies in how integral curves are formed to give train / test data. Namely, for vortex corelines, the dataset contains 25% corelines. However, as a share of the entire volume, the amount of corelines is likely to be much smaller, since it is highly localized to a small number of curves. The latter scenario is more realistic. Moreover, for coreline identification, it is unusual that test frames are interleaved with train frames, yet this is not done for vortex region identification. Common to vortex regions and corelines, length of integration (T) is predefined for all scenarios. However, this would seem to be an important parameter in ensuring integral curves capture vortices - if too small there is insufficient evidence, if too large, curves might describe more than just vortices. This leads to somewhat unfair comparisons with Eulerian methods, whereby velocity grids (or features thereof) are used directly, and no such parameters need be set.

Questions for authors

A broader comment: the paper is rather terse, often conveying important details implicitly in descriptions, if not omitting important rationale.

The paper is motivated by encouraging, though not guaranteeing rotation-invariance. How would the method compare to prior work that employs rotation-equivariant network architectures?

Regarding the latent code used for making vortex predictions, this is taken to be the code associated with the center integral curve. Considering that the center curve only rotates/reflects under action of the symmetry group (it does not permute to match a neighbor, c.f. Eq. 4-left), it is unclear why the symmetry-based invariances enforced for neighbors would impact the representation of the center curve. Parameter-sharing must play a role here, but it would be helpful to clarify this point.

What does "mean coordinates" and "latent mean" refer to (c.f. Appendix C.1)? Is this the coordinate-wise (or equivalently code-wise) mean computed over integration step (T)?

How sensitive is the method to setting the neighbor radius r? The appendix defines this as proportional to the minimum grid spacing, yet it is unclear whether the detection of certain types of vortices would benefit from larger/smaller neighborhoods from which to seed integral curves. The discussion makes reference to FTLE, which is a useful way to think about this, however FTLE does not function as a generic method for vortex extraction.

What is the source of domain shift in Table 6?

1 comment

Thank you for the suggestions. We have clarified the experiments and added comparisons and sensitivity studies.

Motivation beyond conventional extraction methods. Our goal is to identify vortex structures from curve geometry, including particle trajectories for which field-based criteria require an additional velocity-reconstruction step. IVD and VTK provide starting points for reference annotations, which experts inspect and refine. Appendices C.2–C.3 explain this process, while Fig. 5 demonstrates learned clustering directly from particle positions with vorticity shown only for comparison.

IVD thresholds. Figure 2 now displays the expert-selected IVD thresholds in light green, illustrating that suitable thresholds vary substantially across flows. Sensitivity is also flow-dependent: in some flows with a narrow IVD range, even small threshold changes can substantially alter the extent of candidate vortex regions. We will release the dataset and its annotations, including the thresholds.

F1 and Chamfer distance. F1 measures precision and recall for sampled-point classification, whereas Chamfer distance measures geometric discrepancy between extracted and reference corelines. VTK can therefore achieve high F1 while retaining distant spurious segments that substantially increase Chamfer distance. Expert refinement removes these remote false-positive segments, explaining how the two metrics can differ.

Reference to Table 5. We have added an explicit reference to Table 5 in the main text to connect the coreline evaluation discussion with the reported quantitative results.

Coreline sampling proportion. The 25% near-core sampling proportion is a practical dataset-construction choice, not an estimate of the fraction of physical space occupied by corelines. In practice, extraction can focus on candidate vortex regions identified efficiently from quantities such as vorticity. Sampling these regions provides informative positive and negative examples, while unrestricted sampling would introduce many easy background negatives that preliminary filtering could remove.

Interleaved training and test frames. Although flow evolution is temporally continuous, vortex morphology can change rapidly. We interleave training and test frames to cover a broader range of flow states and reduce bias toward structures present only in the initial period. This protocol evaluates recognition across states within a flow; cross-flow generalization is assessed separately on fully unseen flows.

Integration-length sensitivity. Appendix C.6 and Table 15 now evaluate integration length by retraining models at each length while fixing the neighborhood radius and sampled point count. Coreline performance remains relatively stable over length factors of 0.25–1, whereas excessively long integration reduces performance across tasks, consistent with losing the locality of the target structures.

Clarity of experimental details. We have reorganized the appendix to describe curve integration, task-specific readouts and objectives, baseline implementations, and the steps from hairpin-bundle classification to reconstructed surfaces. Table 13 explicitly identifies the training, validation, and test frames.

Comparison with rotation-equivariant networks. We added REViT as a supervised baseline for 2D and 3D vortex-region identification and 3D coreline extraction in Tables 3, 5, 8, and 9. Appendix C.3 describes our voxel-based 3D adaptation using the octahedral symmetry group. In coreline extraction, REViT-3D outperforms Conv3D and PointNet, while StarVortex achieves higher F1, 0.9436 versus 0.9086.

How neighbor constraints affect the center prediction. The prediction is assigned to the center seed, but its latent representation is computed from the entire curve bundle. Shared neighbor features are aggregated with the center-curve information before classification. The symmetry-consistency loss acts on this joint representation, so its gradients shape the neighbor encoding and aggregation used to predict the center label.

Meaning of “latent mean” and “mean coordinates.” “Latent mean” is the encoder-predicted mean vector of the latent distribution. Its coordinates are individual components, not averages over integration steps or spatial positions. We apologize for the ambiguity and have clarified the text.

Neighborhood-radius sensitivity. Appendix C.6 and Table 14 now evaluate six neighborhood radii. The results show that the neighborhood must remain within an appropriate local scale, with performance generally declining at large radii. This study directly evaluates the sensitivity of vortex identification to the neighborhood size.

Source of the domain shift. Table 6 compares both settings on the first five flows in Table 9. Each flow is a domain. For domain generalization, we train on four flows and test on the fifth, averaging all five folds. The shift comes from differences between flows.

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

This paper proposes StarVortex, a method for identifying vortices using the geometry of neighboring particle trajectories and integral curves instead of directly using velocity or vorticity information as neural network inputs. The authors construct groups of curves, called stars, around a central curve and use a convolutional Transformer to learn their representations. They incorporate symmetry constraints and self-supervised objectives to encourage the model to learn geometric features that are consistent under rotations. The authors evaluate their approach on three tasks: vortex-region identification, vortex-coreline extraction, and hairpin-vortex identification. Their results show that the method can identify vortex regions without training labels, while the supervised variants outperform several existing methods on the evaluated tasks.

Critical strengths
  • The paper presents an interesting approach to learning vortex structures directly from the geometry of neighboring trajectories. Unlike many existing methods, the proposed neural network does not require velocity or vorticity values as inputs. The combination of symmetric sampling, rotation consistency, and self-supervised representation learning is also interesting and provides a useful way to learn geometric information from fluid trajectories.
  • The authors evaluate their method on different vortex-related tasks and compare it with several existing approaches. The authors also provide useful ablation experiments and investigate the effects of limited labeled data and domain generalization.
Critical weaknesses
  • One of my main concerns is the generalization of the self-supervised method. In the vortex-region identification experiments, the authors train and cluster the repersentations using the available trajectories from each flow. Therefore, these results do not show whether a model trained on some flow fields can identify vortices in entirely unseen flow fields without retraining. Although the authors include domain generalization experiments for supervised coreline extraction and hairpin identification, they do not provide the same experiment for self-supervised vortex-region identification. I believe this is important because one of the main motivations of the paper is to identify vortices without training labels.
  • Another concern is the practical advantage of using only trajectory geometry. Although the neural network does not require velocity or vorticity as inputs, the trajectories must still be obtained through numerical integration, particle tracking, or interpolation. I could not find an end-to-end computational comparison showing whether this approach is faster or more efficient than existing methods. Also, the results are not consistent across different flow settings. For example, the performance on Channel flow is relatively low for hairpin identification. It is unclear whether these limitations are caused by the available training data, differences between flow fields, or the proposed representation itself.
Questions for authors
  • Can the authors evaluate whether the self-supervised model trained on some flow fields generalizes to entirely unseen flows without retraining? This would help clarify how transferable the learned representations are. Can the authors provide an end-to-end computational comparison that includes trajectory generation, star construction, and neural network inference?
  • What is the main reason for the relatively low performance on Channel flow in hairpin identification? Would additional training data or more diverse flow fields improve the results?
1 comment

Thank you for your constructive feedback. We have expanded the evaluation and clarified the scope of the self-supervised setting.

Self-supervised generalization. We have shown that self-supervised learning objectives can help with domain generalization classification as shown in Table 6, where we tested generalization on five scenes with each of the five scenes held out from both encoder and classifier training. In addition, we did simple cross generalization experiments on the unsupervised 2D vortex identification task on two scenes (Boussinesq to Cylinder2D and Cylinder2D to Boussinesq), and the performance varies case by case. From Boussinesq to Cylinder2D the self-supervised generalization achieves F1 score 0.471 vs. 0.706 trained on Cylinder2D itself; while from Cylinder2D to Boussinesq the generalization is much better with a F1 score of 0.747 vs. 0.760 trained on Boussinesq itself. A more generalizable self-supervised feature encoding system should involve a larger collection of scenes, and that is a research topic out of the scope of this paper and still yet to be explored by future work.

Computational cost. Table 16 now reports curve integration, star construction, prediction, and total extraction time alongside traditional methods. For Re3900, the input consists of stored particle trajectories, so numerical integration is unnecessary. Table 17 additionally compares parameter counts, training times, and estimated test times against deep learning baselines.

Channel-flow performance. Our visual inspection suggests that Channel flow contains many closely adjacent or overlapping hairpins, whose geometry differs substantially from the other two flows. This may contribute to its lower recognition accuracy. Whether more diverse training data would resolve this requires a larger annotated dataset. We therefore present hairpin recognition as an extension application and focus the additional methodological studies on vortex regions and corelines.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebutal Revision

We would like to express our deepest gratitude to all the reviewers for their invaluable time and constructive feedback. To facilitate your review, we have highlighted the most specific modifications in yellow within the provided rebuttal revision PDF

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.