Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Lifting the Winding Number: Precise Discontinuities in Neural Fields for Physics Simulation

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Wind Lifter represents sharp discontinuities in neural fields without storing the cut in the weights.

desk verdict A genuinely new discontinuity-by-lifting construction with sound math; the real question is how far the learned field interpolates in winding-number coordinates. read the letter →

arxiv 2502.00626 v2 pith:KOBRKZAS submitted 2025-02-02 cs.GR

Yue Chang , Mengfei Liu , Zhecheng Wang , Peter Yichen Chen , Eitan Grinspun This is my paper · ORCID
classification cs.GR
keywords neuralfieldsreduced-ordermodelingwindingnumberprogressivecuttingdiscontinuityrepresentationdeformablesimulationcuteditingdiscretization-agnostic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a way to make neural-field bases for reduced-order elastic simulation represent strict displacement discontinuities along a cut curve, without storing the cut's position in the network weights. The central construction is to lift the 2D domain into a 3D graph using the generalized winding number of the cut curve, train an everywhere-continuous volumetric network on that lifted domain, and then evaluate the network on the graph. Because the discontinuity is produced by the restriction, the same trained network supports progressive cutting and interactive edits of the cut geometry—lengthening, moving, or reshaping the cut only requires recomputing the winding-number field analytically. The authors demonstrate this in both data-driven and data-free reduced simulations, reporting real-time frame rates and one-shot generalization to cut placements not seen during training.

What carries the argument

The winding graph $L_\alpha(x) = (x, H_\alpha(x))$, where $H_\alpha$ is the generalized winding number of the portion $\Gamma_\alpha$ of the cut curve that has been made so far and $\alpha$ is the fraction of the curve length cut. The generalized winding number is a harmonic scalar field that jumps by a constant across an open curve, so the graph smoothly separates the two sides of the cut in 3D. Evaluating the continuous network $\tilde{f}_\theta$ on this graph converts an easy learning task—fitting a smooth volumetric field—into a representation of a discontinuous planar field; updating the cut at runtime is just recomputing $H_\alpha$.

What would settle it

Take the trained network for a clockwise helical cut and run the same simulation with a counter-clockwise helical cut, comparing displacement fields against a full-order FEM reference; the paper's Figure 8 shows the deformation diverges in exactly this test. A quantitative version would sweep the gap between training and test winding-number distributions and measure reconstruction MSE, checking whether error grows monotonically with that gap.

Watch

Extended reading notes

Core claim

Restricting a continuous volumetric neural field to the winding-number graph of a cut polyline yields a 2D field with exact, strict discontinuities along that polyline: $f_\alpha(x) = \tilde{f}_\theta(x, H_\alpha(x))$. The key move is to make the jump live in the input coordinates rather than in the network weights, so the network only ever fits a smooth function on $\Omega \times \mathbb{R}$. Since $H_\alpha$ is computed analytically from the current cut curve, the same frozen network can represent a whole family of cut geometries and progressions; the paper demonstrates this by training on one set of cuts and simulating qualitatively different cuts with the weights untouched, and by training the basis without any simulation data using only elastic energy.

Load-bearing premise

The central assumption is that the trained continuous network gives physically correct values on winding-number coordinates it never saw during training; the construction guarantees a sharp discontinuity for any cut, but not that the deformation on an out-of-distribution cut is the right one.

Editorial extensions

If this is right

  • A cut can be lengthened, moved, or reshaped during a simulation by re-evaluating the winding graph, with no retraining and no remeshing.
  • A single trained network generalizes to cut placements not present in the training distribution, as demonstrated on one-shot tests and interactive editing.
  • Data-free training is possible: minimizing elastic energy alone produces a cut-aware basis, which prior mesh-free data-free methods could not do.
  • Reduced simulation with cut changes runs at interactive rates (roughly 28–42 fps with cut updates, 62–166 fps without), giving a 30×–183× speedup over the full-space solver used for comparison.
  • The representation also generalizes to external forces unseen during training, and the simulated sagging of a cut paper helix qualitatively matches a physical experiment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The lifting construction is not tied to winding numbers in principle; any analytically computable jump function would separate the sides of a cut, so the same idea may transfer to discontinuities in other settings, such as fluid interface jumps or 3D crack surfaces.
  • The paper's own out-of-distribution test suggests the practical limit is interpolation in winding-number space, not the sharpness of the discontinuity; this points to a testable extension where training on a family of winding graphs rather than one would widen the range of editable cuts.
  • Because cut geometry is an explicit polyline, the representation could be plugged into derivative-based design tools, where optimizing a cut shape for a target deformation becomes an optimization over a few control points with the neural basis held fixed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces "Wind Lifter," a method for representing fields with strict discontinuities along curves using continuous neural fields. The key idea (Section 3, Eq. 3) is to append a generalized winding number H_α(x) as a third input coordinate, forming a graph L_α(x)=(x, H_α(x)) in 3D, and to define the 2D field as the restriction f = f̃_θ ∘ L_α of a learned continuous volumetric field to this graph. Because H_α is discontinuous across the cut Γ_α, the restriction is strictly discontinuous there, while the neural network itself remains continuous. The authors apply this construction to reduced-order modeling of thin shells, training the volumetric field either data-driven (Eq. 8) or data-free by minimizing elastic energy (Eq. 9). They demonstrate progressive cutting, runtime editing of cut polylines, generalization to held-out cut shapes and unseen loads, and interactive design, and compare against LiCROM, DANN, and Simplicits baselines (Figure 14).

Significance. The central mathematical construction is elegant and likely correct: it cleanly decouples the geometry of the discontinuity (encoded in the analytically computed winding graph) from the learnable continuous part, so the location of the jump is exact by construction without re-training. The paper also makes a useful contribution by showing that lifting is compatible with both data-driven and data-free training of reduced bases, and it demonstrates real-time interactive cut editing, which is not available in prior mesh-based or weight-encoded neural ROMs. The empirical results on the demonstrated examples are promising and show clear improvements over the baselines. If the generalization behavior is confirmed by a broader quantitative study, the work would be a substantive advance in reduced-order modeling of cutting. The paper is honest about its limitation (Section 6, Out-of-Distribution Challenges), though the abstract and contributions state the generalization claim more strongly than the evidence currently supports.

major comments (3)
  1. [Section 5.2, Figure 14; Section 6] The claim of generalization to "cut placements not seen during training" (Section 1) rests on exactly one held-out test shape (Figure 14, MSE=0.20% versus 0.02% on the training shape) and on qualitative results (Figures 7 and 8). The paper's own Section 6 acknowledges that performance decreases "when tested on winding number distributions significantly different from the training data," and Figure 8 shows that a counter-clockwise helix yields a deformation different from the training-case result. This does not invalidate the construction—Eq. (3) guarantees the location and existence of a strict jump for any Γ_α—but it means the central generalization claim is currently supported only in an unquantified neighborhood of the training distribution. I request one of the following: (i) a systematic error study across multiple random held-out cut polylines with varying geometric deviation, including error bars over network initializations; (ii) an explicit characterization of the region in cut-parameter space (or winding-number-coordinate space) in which the reconstruction error is below a stated tolerance; or (iii) a training augmentation, such as superimposing the loss on multiple winding graphs per epoch, that demonstrably covers the intended test distribution.
  2. [Section 3, Eq. (3); Section 4.1, Eqs. (8)–(9)] The construction f = f̃_θ ∘ L_α makes the location and existence of the discontinuity rigorous, but it makes no statement about the correctness of the field values on lifted coordinates that were absent from training. The losses in Eqs. (8) and (9) are evaluated only on the winding graphs of the training cuts; for a test cut Γ_test, the third coordinate H_test(x) can lie outside the range of H seen in training, and the loss does not constrain those regions. Therefore, both the magnitude of the jump and the deformed shape on an unseen cut depend on interpolation behavior of the trained volumetric network, not on the lifting construction. This is a gap between the mathematical statement and the empirical generalization claim, not a circularity: the held-out test in Figure 14 does produce nontrivial errors (0.20% versus 0.02%), which shows the test is not a restatement of fitted values. To make the claim load-bearing, the revision should either characterize the interpolation error on out-of-distribution winding-number coordinates (e.g., by plotting the training and test distributions of the lifted coordinate z and reporting errors as a function of distribution distance), or train with a distribution of winding graphs broad enough to cover the editing and generalization scenarios advertised in Sections 1 and 5.
  3. [Section 5.2, Figure 14; Table 1] The quantitative comparison reports a single MSE value per method and condition, with no variance across training runs or across multiple test cuts. Because the headline contribution is generalization to unseen cuts, the evaluation should report mean and standard deviation over at least three to five independent network initializations and over several randomly generated held-out cut shapes. As written, it is impossible to tell whether the 0.20% test error for the single unseen cut is representative or a favorable draw, particularly since the same figure shows the baselines at 2.10%–2.74%.
minor comments (5)
  1. [Section 4.3, Strain singularity at crack tip; Section 5] The smoothing kernel width ε and the winding-number scaling factor of 32 are free parameters; please state how they were chosen and, if possible, include a brief sensitivity study to show that the reported results are not sensitive to their exact values.
  2. [Section 4.1, Eq. (9)] Equation (9) has a punctuation issue: the equation is followed by a comma before "where." Also, the notation L_data-free = E_elas is not written as an explicit function of the network parameters θ and the cut parameter α; clarifying this would help the reader.
  3. [Figure 12] The y-axis label "MSE (%)" is ambiguous because the axis ranges from about 10^-1 to 10^0; please specify whether these are percentages and what the normalization is (for example, relative to what reference deformation).
  4. [Section 5.2, Figure 14] The baseline "DANN" is named in Figure 14 but not described in the text beyond the citation to Belhe et al. [2023]; add a one-sentence description of the method so the comparison is self-contained.
  5. [Section 5.1, Figure 7] The data-free generalization test with three held-out cut shapes is presented only qualitatively; adding quantitative errors for these cases would help support the generalization claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the lifting construction is a definitional mechanism; held-out cut tests and the paper's own OOD caveat confirm the generalization claim is not a restatement of the training fit.

full rationale

The Wind Lifter construction (Eq. 3: f = f_tilde_theta ∘ L) is an explicit definitional device, not a fitted input masquerading as a prediction: the discontinuity location is inserted analytically via H_alpha (Eqs. 4-5), and the network only learns the volumetric field f_tilde_theta. The central generalization claim is supported by held-out experiments (Figure 7; Figure 14 bottom: MSE 0.20% on a test cut vs 0.02% on a training cut), which are genuine evaluations across different cut geometries, so the result does not reduce to the training fit. Self-citations to LiCROM [Chang et al. 2023] for the training loss (Eq. 8), the subspace update (Eq. 10), and the data-free baseline [Modi et al. 2024] (Eq. 9) are methodological reuse; they are not used to establish the lifting construction or to forbid alternative representations, so they are not load-bearing. The paper's own Section 6 out-of-distribution caveat (Figure 8: 'our method's performance decreases when tested on winding number distributions significantly different from the training data') is a genuine limitation on extrapolation, not circularity; if anything, the observed degradation confirms that the held-out tests are not tautological. No step in the derivation chain is equivalent by construction to its inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 2 invented entities

The construction introduces a new input dimension (the lifted winding number) and conditions the neural field on cut extent α, but no new physical entity or force is posited. The central claim rests on standard properties of generalized winding numbers, the standard linear-subspace ROM ansatz, and an empirical interpolation assumption that the network generalizes across winding-number distributions. Hand-chosen parameters include the winding scaling factor, basis size, network architecture, and crack-tip smoothing width. The paper's own limitations (Figure 8, Section 6) confirm that the out-of-distribution interpolation assumption is load-bearing.

free parameters (4)
  • Winding number scaling factor = 32
    Section 5 states that winding numbers are scaled by a factor of 32 to ensure sufficient separation of lifted and 2D coordinates. This hand-chosen scalar changes the geometry of the lifted input space and affects training and generalization.
  • Crack-tip smoothing kernel width ε = not specified in main text
    Section 4.3 applies a cubic spline kernel in an ε-neighborhood of cut endpoints to regularize the winding-number gradient; the kernel and ε are deferred to the supplementary material. It alters the near-tip displacement gradient, so it is a hand-chosen free parameter.
  • Number of reduced basis modes k = 18 (data-free), 20 (data-driven)
    Section 5.1 uses an 18-dimensional basis for data-free examples; Section 5.2 trains k=20 modes for data-driven examples. The subspace dimension is selected by the authors, not derived, and bounds what deformations the reduced simulation can represent.
  • Neural network architecture hyperparameters = 5-layer, 128-channel SIREN, positional encoding up to frequency 23
    Section 5 lists these as the default settings. They control the smoothness and capacity of the volumetric field and are chosen by hand; no systematic ablation is reported beyond Figure 13's activation and scale comparison.
assumptions (6)
  • standard math The generalized winding number H_α(x) from Eq. 5 is a well-defined analytic function on Ω∖Γ_α with a fixed jump across the cut line Γ_α.
    Used in Section 3 to define the lifting operator L_α and in Eq. 7 for the restricted field. This relies on properties established in Jacobson et al. 2013 and Barill et al. 2018, which the paper cites.
  • domain assumption Displacement is a linear combination of k reduced modes, u = z^T Φ_α, and the reduced configuration z evolves by the optimization in Eq. 10.
    This is the ROM ansatz introduced in Section 4. It is load-bearing for the real-time speed claim; if deformation complexity exceeds k modes, the reduced simulation will be inaccurate.
  • domain assumption The elastic energy Ψ is a hyperelastic density (StVK in data-free, stable Neo-Hookean in data-driven) evaluated with deformation gradients from automatic differentiation of the neural basis.
    Used in Eqs. 9 and 10 for data-free training and dynamic simulation. Standard continuum mechanics, stated in Sections 4.1 and 4.2, but the specific material models are not the paper's contribution.
  • ad hoc to paper For runtime edits and unseen cut shapes, the network must interpolate accurately on winding-number values outside its training distribution.
    The generalization claim depends on this empirical assumption. Section 6 and Figure 8 acknowledge that significant distribution shift degrades results, so this is a load-bearing premise rather than a derived property.
  • domain assumption Cut geometry is limited to piecewise-linear polylines, and progressive cutting is parameterized by the first α-fraction of polyline arc length.
    Section 4.3 describes cut representation as a collection of polylines and integration bounds advanced by walking along the polyline. This excludes curved cuts and arbitrary cut evolution, an explicit limitation in Section 6.
  • ad hoc to paper The crack-tip strain singularity can be smoothed by multiplying the winding number with a cubic spline in an ε-neighborhood without materially changing the simulated deformation.
    Introduced in Section 4.3 to handle the divergent gradient of H near cut endpoints. No convergence study or accuracy analysis of ε is given, so the physical validity of the regularization is assumed.
invented entities (2)
  • Lifted winding-number coordinate z = H_α(x)
    purpose: Extra input dimension to the neural field that separates the two sides of a cut, turning the problem of learning discontinuous fields into learning a continuous volumetric field.
    Core construction of Section 3, Eqs. 4 to 7. It is an input feature, not a physical entity; it has no falsifiable prediction outside the paper's own experiments.
  • Winding graph
    purpose: The graph of H_α over Ω, used as the restriction manifold so that a continuous neural field becomes a discontinuous planar field after restriction.
    Introduced in Section 3: 'We refer to the graph of H(x) as the winding graph.' It is a geometric visualization and construction device, not independently testable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lifting the Winding Number: Precise Discontinuities in Neural Fields for Physics Simulation." pith.science (2026). https://pith.science/paper/KOBRKZAS

@misc{pith2026250200626,
  author       = {Pith},
  title        = {Pith review of: Lifting the Winding Number: Precise Discontinuities in Neural Fields for Physics Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KOBRKZAS}},
  note         = {Machine review of arXiv:2502.00626}
}
read the original abstract

Cutting thin-walled deformable structures is common in daily life, but poses significant challenges for simulation due to the introduced spatial discontinuities. Traditional methods rely on mesh-based domain representations, which require frequent remeshing and refinement to accurately capture evolving discontinuities. These challenges are further compounded in reduced-space simulations, where the basis functions are inherently geometry- and mesh-dependent, making it difficult or even impossible for the basis to represent the diverse family of discontinuities introduced by cuts. Recent advances in representing basis functions with neural fields offer a promising alternative, leveraging their discretization-agnostic nature to represent deformations across varying geometries. However, the inherent continuity of neural fields is an obstruction to generalization, particularly if discontinuities are encoded in neural network weights. We present Wind Lifter, a novel neural representation designed to accurately model complex cuts in thin-walled deformable structures. Our approach constructs neural fields that reproduce discontinuities precisely at specified locations, without baking in the position of the cut line. Crucially, our approach does not embed the discontinuity in the neural network's weights, opening avenues to generalization of cut placement. Our method achieves real-time simulation speeds and supports dynamic updates to cut line geometry during the simulation. Moreover, the explicit representation of discontinuities makes our neural field intuitive to control and edit, offering a significant advantage over traditional neural fields, where discontinuities are embedded within the network's weights, and enabling new applications that rely on general cut placement.

Figures

Figures reproduced from arXiv: 2502.00626 by the authors.

Figure 1
Figure 1. Our method calculates a basis for reduced-space simulations that can represent shapes with a family of discontinuities. This basis is represented by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We augment the input to our neural field with a generalized winding number, lifting the domain from 2D into 3D. As the cut progresses, the generalized [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Our method is able to capture the complex cut of this butterfly shape, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The cut change can lead to a big difference in discretization using [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Our method enables interactive drawing and editing of cut shapes [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 10
Figure 10. Figure 10: We trained the basis on simulations of tugging on a [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: We can interactively edit the cut shape for the kirigami. The cut [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 7
Figure 7. Figure 7: We train our method with the three blue cut shapes (top), and test it [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 13
Figure 13. Figure 13: We compared the percentage mean squared error (MSE) of our method with LiCROM [Chang et al. 2023] using different MLP scales and activation functions. Our method achieves significantly lower error across all settings, as shown on the left. Additionally, it demonstrate…
Figure 14
Figure 14. Figure 14: We have compared our percentage reconstruction error with other methods. Our method has the lowest reconstruction error for both shapes in and not in the training set. Moreover, our method is the only one that can generate a visually reasonable result when tested on a…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Precise Gradient Discontinuities in Neural Fields for Subspace Physics

    cs.GR 2025-05 conditional novelty 6.0 of 10

    Augmenting neural field inputs with a clamped distance function lets a single network represent gradient discontinuities at material interfaces and creases, enabling discretization-agnostic reduced-order physics simulation.

Reference graph

Works this paper leans on

7 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [2009]

    Graph.28, 3, Article 50 (jul 2009), 10 pages

    Enrichment textures for detailed cutting of shells.ACM Trans. Graph.28, 3, Article 50 (jul 2009), 10 pages. https://doi.org/10.1145/1531326.1531356 Byungsoo Kim, Vinicius C Azevedo, Nils Thuerey, Theodore Kim, Markus Gross, and Barbara Solenthaler. 2019. Deep fluids: A generative network for parameterized fluid simulations. InComputer Graphics Forum, Vol....

    arXiv 2024

  2. [2016]

    InAcm siggraph 2016 courses

    The material point method for simulating continuum materials. InAcm siggraph 2016 courses. 1–52. Peter Kaufmann, Sebastian Martin, Mario Botsch, Eitan Grinspun, and Markus Gross

    work page 2016

  3. [2018]

    Graph.37, 4, Article 60 (jul 2018), 14 pages

    Tetrahedral meshing in the wild.ACM Trans. Graph.37, 4, Article 60 (jul 2018), 14 pages. https://doi.org/10.1145/3197517.3201353 Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. 2013. Robust inside-outside segmentation using generalized winding numbers.ACM Trans. Graph.32, 4, Article 33 (jul 2013), 12 pages. https://doi.org/10.1145/2461912.2461916...

    arXiv 2018

  4. [2019]

    David Grangier, Pierre Ablin, and Awni Hannun

    Latent-space Dynamics for Reduced Deformable Simulation.Computer Graphics Forum(2019). David Grangier, Pierre Ablin, and Awni Hannun. 2023. Adaptive training distributions with scalable online bilevel optimization.arXiv preprint arXiv:2311.11973(2023). Yixin Hu, Qingnan Zhou, Xifeng Gao, Alec Jacobson, Denis Zorin, and Daniele Panozzo

    arXiv 2019

  5. [2020]

    kirigami tower

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. InAdvances in Neural Information Processing Systems, Vol. 33. 7537–7547. Yuanyuan Tao, Ivan Puhachov, Derek Nowrouzezahrai, and Paul Kry. 2024. Neural Implicit Reduced Fluid Simulation. InSIGGRAPH Asia 2024 Conference Papers. 1–11. Quoc-Minh Ton-That, Paul G. Kry, and...

    arXiv 2024

  6. [2021]

    InProceedings IEEE/CVF Conf

    On Self-Contact and Human Pose. InProceedings IEEE/CVF Conf. on Computer Vision and Pattern Recogßnition (CVPR). Octavi Obiols-Sales, Abhinav Vishnu, Nicholas Malaya, and Aparna Chandramowlish- waran. 2023. ADARNet: Deep Learning Predicts Adaptive Mesh Refinement. In Proceedings of the 52nd International Conference on Parallel Processing. 524–534. Shaowu ...

    arXiv 2022

  7. [2024]

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger

    Accelerate Neural Subspace-Based Reduced-Order Solver of Deformable Simulation by Lipschitz Optimization.ACM Transactions on Graphics (TOG)43, 6 (2024), 1–10. Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. 2019. Occupancy networks: Learning 3d reconstruction in function space. InProceedings of the IEEE/CVF confer...

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.