Pith. sign in

REVIEW 4 major objections 5 minor 44 references

PDEfuncta: Spectrally-Aware Neural Representation for PDE Solution Modeling

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A spectrally reparameterized modulation lets one shared neural network encode many high-frequency PDE solution fields and supports forward and inverse inference from the same latent codes.

desk verdict A simple and effective modulation trick for Functa-style PDE compression; the empirical claims are strong, but the spectral-bias theory is borrowed, not re-derived, and the generalization story is thin. read the letter →

arxiv 2506.12790 v1 pith:PGCCB3NG submitted 2025-06-15 cs.LG cs.NAmath.NAphysics.comp-ph

Minju Jo , Woojin Cho , Uvini Balasuriya Mudiyanselage , Seungjun Lee , Noseong Park , Kookjin Lee This is my paper
classification cs.LGcs.NAmath.NAphysics.comp-ph
keywords implicitneuralrepresentationsspectralbiasFourierreparameterizationPDEsolutionfieldsmodulatedINRsoperatorsbidirectionalinferencefuncta
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

The paper claims that spectral bias in shared implicit neural representations can be overcome by reparameterizing every layer weight with a fixed Fourier basis and an additive per-sample modulation, producing compact latent codes that reconstruct high-frequency PDE solution fields. It introduces PDEfuncta, which uses one shared latent vector to modulate a pair of INRs representing two related function spaces, so that forward and inverse mappings between them can be performed by reading off the other field from the latent. If true, this gives a single lightweight representation that compresses large scientific datasets, generalizes to unseen PDE parameters via latent interpolation, and matches or beats dedicated neural operators on geometry benchmarks.

What carries the argument

The mechanism is the layer-wise weight construction $W_k = S_k \Phi_k = (R_k + \alpha_{k,W}\cdot 1^\top)\Phi_k$, where $\Phi_k$ is a fixed Fourier basis whose rows are phase-shifted cosines at hand-picked low and high frequencies, $R_k$ is a shared base coefficient matrix across all samples, and $\alpha_{k,W}$ is the per-sample modulation produced by a small network from the latent $z$. This places each sample's weights in a predefined frequency subspace, giving better neural tangent kernel conditioning and stronger high-frequency gradients (per Theorem 1 quoted from the Fourier reparameterization literature), while keeping instance-specific adaptation cheap and letting the same $z$ drive two networks that jointly represent paired spaces.

What would settle it

Run GFM on a solution dataset whose dominant spatial frequencies fall well outside the configured $n_{\text{low}}$–$n_{\text{high}}$ band and check whether its PSNR lead over FiLM collapses; alternatively, check directly whether the Theorem 1 gradient-ratio inequality still holds when the coefficient matrix is constrained to $R + \alpha 1^\top$ rather than being free.

Watch

Extended reading notes

Core claim

The paper's central claim is that Global Fourier Modulation (GFM), the weight reparameterization $W_k = (R_k + \alpha_{k,W} 1^\top)\Phi_k$ with a fixed cosine basis $\Phi_k$, removes the spectral-bias bottleneck of modulated INRs: high-frequency modes receive stronger gradient signals than in plain or activation-space-modulated networks, so one shared backbone and a 20-dimensional latent vector can accurately encode many solution fields at once. Building on that, PDEfuncta ties two such INRs (for spaces $A$ and $U$) to the same latent vector, so a single code represents both fields and new tasks, forward or inverse, are handled by latent inference without retraining the shared weights. The paper supports this with reconstruction comparisons against Shift, Scale, and FiLM across convection, Helmholtz, Navier–Stokes, and Kuramoto–Sivashinsky data, generalization experiments on unseen convection coefficients, and operator-learning comparisons on airfoil and pipe geometry benchmarks.

Load-bearing premise

The paper's advantage rests on the fixed Fourier basis, with frequencies chosen by hand per dataset, actually covering the spectrum of every new field; the authors themselves note that performance depends on tuning the frequency ratio.

Editorial extensions

If this is right

  • A single 20-dimensional latent code can compress and reconstruct multiple high-frequency PDE solution fields that Shift, Scale, and FiLM cannot represent at comparable fidelity.
  • Unseen PDE coefficients become reachable by interpolating or fitting latent codes, so inference on novel parameter regimes does not require retraining the shared network.
  • A shared latent vector across paired spaces yields bidirectional operators $A \to U$ and $U \to A$, so forward and inverse problems are solved by the same representation.
  • The frequency-axis design gives the practitioner explicit control over the low-frequency and high-frequency mix, which can be tuned per dataset.

Reading between the lines

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

  • The hand-picked frequencies ($n_{\text{low}}$, $n_{\text{high}}$, $n_{\text{phase}}$) are the main transfer bottleneck: if the paper's recipe is applied to datasets whose spectra are unknown, automatic frequency selection would be the natural next step.
  • Because the shared basis weights and modulations are differentiable, the same latent could in principle be optimized jointly with a physics loss, turning PDEfuncta into a PINN-style solver to speed inference further; the authors do not test this.
  • The claimed advantage should be read as conditional on the SIREN backbone used throughout; other activations may interact differently with the Fourier basis.
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

4 major / 5 minor

Summary. The paper proposes Global Fourier Modulation (GFM), a modulation technique for implicit neural representations in which each layer's weight matrix is written as W_k = (R_k + alpha_{k,W} 1^T) * Phi_k, with a shared base matrix R_k, a fixed Fourier basis Phi_k, and a per-sample modulation vector alpha_k. Building on GFM, the authors introduce PDEfuncta, a Functa-style framework that encodes paired input/output solution fields with a single shared latent vector and supports bidirectional inference between the two function spaces. The empirical sections report consistent PSNR/MSE improvements over Shift, Scale, and FiLM on five PDE reconstruction benchmarks, generalization to unseen convection coefficients through latent interpolation and partial-observation fitting, and competitive or better L2 errors than several neural-operator baselines on airfoil and pipe-flow benchmarks. The theoretical discussion invokes an existing Fourier-reparameterization theorem to argue that GFM mitigates spectral bias, although the theorem is stated for a free coefficient matrix rather than GFM's constrained additive form.

Significance. If the empirical claims hold, GFM is a simple and practical modulation mechanism that improves the spectral fidelity of dataset-level INRs at modest computational overhead, and the bidirectional Functa formulation is a useful idea for compression and inverse problems in scientific machine learning. The paper's strengths include direct ablations that isolate the modulation mechanism, standard deviations reported for the main reconstruction results in Appendix F.3, clean generalization experiments on unseen PDE coefficients, and computational cost measurements in Appendix E.2. The main weaknesses are that the theoretical justification does not cover the constrained parameterization actually used, the neural-operator comparison leaves the test-time latent-inference protocol unspecified, and the paper does not report parameter-count or capacity-matched comparisons against its baselines.

major comments (4)
  1. [Section 3.2, Eq. (3) and Eq. (8)] The displayed dimensions in Eq. (3) are inconsistent with the rest of the paper. Section 3.1 defines Phi_k in R^{D x M}, and Section 4.1 states that the additive modulation is applied to a matrix in R^{M x D}, but Eq. (3) writes R_k in R^{d x M} and multiplies (R_k + alpha_{k,W} 1^T) in R^{d x M} by Phi_k in R^{D x M}, which is undefined unless Phi_k is transposed or the dimensions are changed. Eq. (8) repeats the same product. Since this parameterization is the paper's central mechanism, the intended dimensions must be stated unambiguously and matched to the basis construction in Appendix H, where D = (n_low + n_high) * n_phase.
  2. [Section 3.2, Theorem 1] The statement that GFM 'inherits the benefits of Fourier reparameterization—such as improved NTK conditioning and stronger high-frequency response (Theorem 1)' is not supported as written. Theorem 1 is an existence result for a free trainable coefficient matrix S in R^{M x D}, whereas GFM uses the constrained family S_k = R_k + alpha_{k,W} 1^T, which is a shared base matrix plus a rank-one modulation term per layer. The theorem's gradient-ratio bound does not automatically transfer to this low-dimensional affine subspace, and the paper does not re-derive it for the constrained case. The limitation is compounded by the paper's own acknowledgement in Appendix B that optimal performance depends on tuning the frequency ratio (n_low, n_high, n_phase). Please either prove an analogue of the theorem for the constrained parameterization or explicitly restate the theoretical discussion as a heuristic motivation rather than a guarantee.
  3. [Section 5.3.2, Table 4] The neural-operator evaluation protocol for unseen samples is not specified. A standard operator baseline such as FNO receives the input field A and predicts U in one forward pass, but PDEfuncta must obtain a latent code z for each test sample before it can decode either field. If z is found by test-time gradient fitting against the full target field U, or against A plus some other information, the comparison with one-pass neural operators is not operationally equivalent. Additionally, the text states that the authors 'evaluate both G and G† on unseen samples,' but Table 4 reports only the forward mapping G, with no numerical inverse-mapping errors. Please specify which modality is observed at test time, how z is obtained, how many inner-loop steps are used, and include the inverse-mapping results in the table.
  4. [Section 5.2, Table 1 and Appendix E.2] No parameter-count or capacity-matched comparison is reported against the modulation baselines. GFM's coefficient matrix S_k has D columns per layer, which can be substantially larger than the standard weight matrices used by Shift, Scale, and FiLM; for example, the Helmholtz setting uses n_low = n_high = 128 and n_phase = 32, giving D = 8192 basis vectors per layer. The gains in Table 1 could therefore partly reflect additional model capacity rather than the spectral modulation mechanism. Table 6 reports memory and time, but not parameter counts. Please report parameter counts for all modulation methods and include a capacity-matched ablation, for instance by increasing the width or depth of the baseline backbones.
minor comments (5)
  1. [Appendix E.1] There are several typos and grammatical issues, including 'bach size' instead of 'batch size' in Appendix E.1, 'model struggle' in Section 3.1, 'Functation space mapping' in Table 5, and 'resolution free characteristic' in the Conclusion.
  2. [Section 3.1, Eq. (1)] The gradient ratio in Eq. (1) is written without norms, so the expression is ambiguous; it should be stated with norms, e.g., |dL(F1)/ds_{i,m}| / |dL(F2)/ds_{i,m}|, or with a precise definition of the quotient.
  3. [Appendix H] Appendix H refers to 'Equation 3.1' when it means Eq. (2), and the definition T_max = 2*pi*n_low appears inconsistent with the statement in Section 3.1 that T_max is determined by the smallest frequency used; please reconcile the two descriptions.
  4. [Appendix J, Figure 16] The caption of Figure 16 says the results are for 'the convection equation,' but the displayed data and surrounding text concern the Helmholtz equation with coefficients (a1, a2); the caption should be corrected.
  5. [Figure 1] Figure 1 is referenced in the introduction but its axes and color scale are not described, and the caption does not explain how the reader should see the spectral bias from the panel; please add a clearer caption or axis labels.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity found: the central claims rest on genuinely out-of-sample tasks and external benchmarks; the only self-citation is a non-load-bearing data attribution, and the imported theorem is from non-author prior work.

full rationale

No significant circularity found; every load-bearing empirical claim is tested genuinely out-of-sample, and the only imported theoretical result comes from external prior work. GFM's weight construction W_k = (R_k + alpha_k,W * 1^T) * Phi_k (Eq. 3) is a constrained instance of the external Fourier reparameterization W = S * Phi from [38], so the quoted Theorem 1 (Shi, Zhou, Gu, CVPR 2024, non-author prior work) is formally applicable in shape; the paper does not re-derive the gradient bound for the additive form nor for the multi-sample meta-learning loop, which is an incompleteness of the theoretical framing, not a circular step, since the theorem's assumptions do not include the empirical results it is used to motivate. The unseen-coefficient studies (Section 5.2.1) are genuine extrapolation tests: in Setting 1 the latent for beta* is obtained by cubic interpolation of training latents with no access to the target solution, and in Setting 2 only the partial segment t in [0,0.5] is used to fit z while accuracy on the unobserved t in [0.5,1] is the reported generalization. The operator-learning results (Table 4) are evaluated against external baselines (FNO, UNet, Geo-FNO, FFNO, CORAL, MARBLE) on held-out Euler-NACA and Pipe geometries. The sole self-citation, [44] (co-author Seungjun Lee), appears only as a provenance citation for Navier-Stokes data in Appendix D.4 and is not load-bearing. The paper itself flags its own main limitation in Appendix B ('optimal performance may require careful hyperparameter tuning to select the appropriate frequency ratio'), and Appendix H confirms the frequencies are hyperparameters; this heuristic basis selection is a robustness risk for unseen spectra, but Table 1's advantage is established by direct ablation against Shift, Scale, and FiLM at matched latent dimension (20) and comparable memory, so the comparisons do not presuppose their conclusion.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on hand-tuned spectral hyperparameters (n_low, n_high, n_phase), a borrowed theorem that is not re-proved for the constrained modulation, and the assumption that a single latent can jointly encode paired fields. No new physical entities are introduced.

free parameters (4)
  • Fourier basis frequency counts (n_low, n_high, n_phase) = n_phase=32 fixed; n_high/n_low: Convection 128/32, Helmholtz 128/128, KS 64/16, NS 8/128, FWI 128/32
    Chosen by hand per dataset in Appendix E.1 and Appendix H; these determine the spectral content of the fixed basis and are acknowledged in Appendix B to require tuning.
  • Latent code dimension = 20
    Fixed to 20 for all experiments in Appendix E.1; a capacity choice shared across methods, but it affects the expressiveness of modulation.
  • Inner and outer learning rates = eta_inner=0.01, eta_outer=0.0001
    Set globally in Appendix E.1; standard meta-learning hyperparameters.
  • Backbone architecture = SIREN, 5 layers, hidden 256, modulation MLP hidden 512
    Fixed for all methods in Appendix E.1; architecture choice affects capacity and could interact with the modulation method.
assumptions (3)
  • ad hoc to paper Theorem 2 of Shi et al. [38] applies to GFM's constrained coefficient matrix S = R + alpha 1^T even though it is proved for a free trainable S.
    Invoked in Section 3.2 to claim GFM inherits spectral-bias mitigation; no proof is offered for the additive modulation constraint.
  • domain assumption The fixed Fourier basis with n_low, n_high, and n_phase frequencies provides sufficient spectral coverage for every target PDE field.
    Basis construction in Appendix H is heuristic; Appendix B notes performance depends on tuning the frequency ratio.
  • domain assumption A single 20-dimensional latent vector can jointly encode both sides of a paired function space so that bidirectional operator inference is meaningful.
    Core to PDEfuncta in Section 4; demonstrated only on a limited set of datasets and without theoretical analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PDEfuncta: Spectrally-Aware Neural Representation for PDE Solution Modeling." pith.science (2026). https://pith.science/paper/PGCCB3NG

@misc{pith2026250612790,
  author       = {Pith},
  title        = {Pith review of: PDEfuncta: Spectrally-Aware Neural Representation for PDE Solution Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PGCCB3NG}},
  note         = {Machine review of arXiv:2506.12790}
}
read the original abstract

Scientific machine learning often involves representing complex solution fields that exhibit high-frequency features such as sharp transitions, fine-scale oscillations, and localized structures. While implicit neural representations (INRs) have shown promise for continuous function modeling, capturing such high-frequency behavior remains a challenge-especially when modeling multiple solution fields with a shared network. Prior work addressing spectral bias in INRs has primarily focused on single-instance settings, limiting scalability and generalization. In this work, we propose Global Fourier Modulation (GFM), a novel modulation technique that injects high-frequency information at each layer of the INR through Fourier-based reparameterization. This enables compact and accurate representation of multiple solution fields using low-dimensional latent vectors. Building upon GFM, we introduce PDEfuncta, a meta-learning framework designed to learn multi-modal solution fields and support generalization to new tasks. Through empirical studies on diverse scientific problems, we demonstrate that our method not only improves representational quality but also shows potential for forward and inverse inference tasks without the need for retraining.

Figures

Figures reproduced from arXiv: 2506.12790 by the authors.

Figure 1
Figure 1. [Spectral bias problem] Experimental re [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of GFM. The GFM framework injects fixed Fourier bases into each layer of the shared INR, enabling effective learning of both low- and high-frequency components. Theorem 1 from [11] provides a theoretical explanation for this phenomenon, showing that low￾frequency components dominate the gradient dynamics near initialization. As a result, model struggle to capture fine-scale structures, which are ess… view at source ↗
Figure 3
Figure 3. Proposed method: PDEfuncta. PDEfuncta leverages GFM to represent paired function spaces using a shared latent vector, supporting bidirectional inference between input and output fields. This architecture enables efficient compression and reversible mapping for scientific datasets. FiLM—operate solely in the activation space and lack explicit frequency control. This limitation makes them inadequate for modeling signa… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: [Kuramoto–Sivashinsky] Training loss (MSE) over epochs. We begin with the standard single-INR setting, where a single INR is modulated by a sample-specific latent codes to com￾press and reconstruct individual continuous solution fields. Ex￾periments are conducted on fo…
Figure 6
Figure 6. Figure 6: Top: Error maps for the Kuramoto–Sivashinsky equation. Bottom: Helmholtz equation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: [Convection] Reconstruction results for unseen β = 24.5, 49.5 on Setting 1 and 2. Our experimental results, which are summarized in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: PDEfuncta reconstructs both forward G † : A → U and inverse G : U → A mappings between function spaces for unseen airfoil and pipe flow samples. 5.3.1 Reconstruction Fidelity on Seen Samples: Modulation Comparison We evaluate PDEfuncta’s ability to compress and reconst…
Figure 8
Figure 8. Figure 8: Training loss (MSE) curves for Shift, Scale, FiLM, and GFM on (a) Convection, (b) [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Dynamics of Latents (convection equations) [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: [Convection] Predicted solution snapshots for unseen coefficient β = 30.5 from interpolated latent codes using different modulation methods. (a) Ground truth, (b) Shift, (c) Scale, (d) FiLM, and (e) GFM. For each method, the latent code for β = 30.5 is obtained via cu…
Figure 11
Figure 11. Figure 11: Visualization of the PDEfuncta pipeline for the Airfoil dataset. Two modalities in airfoil data are represented as separate INRs fa and fu, and are jointly compressed and reconstructed through a shared latent modulation vector z. This section provides supplementary ex…
Figure 12
Figure 12. Figure 12: Ground truth for a representative FWI sample. (a) Velocity map and (b–f) correspond￾ing seismic data. (a) Velocity (b) Seismic #1 (c) Seismic #2 (d) Seismic #3 (e) Seismic #4 (f) Seismic #5 (g) Velocity (h) Seismic #1 (i) Seismic #2 (j) Seismic #3 (k) Seismic #4 (l) S…
Figure 13
Figure 13. Figure 13: FWI bidirectional reconstruction results for seen samples using different modulation methods. Each row shows a velocity map and its corresponding seismic data reconstructed by a different modulation method: (a–f) Shift, (g–l) Scale, (m–r) FiLM, and (s–x) GFM. 20 [PIT…
Figure 14
Figure 14. Figure 14: Train loss (MSE) curves comparing Spatial Functa and GFM across epochs for PDE [PITH_FULL_IMAGE:figures/full_fig_p022_14.png]
Figure 15
Figure 15. Figure 15: Reconstruction results for Spatial Functa and GFM on the convection equation with [PITH_FULL_IMAGE:figures/full_fig_p023_15.png]
Figure 16
Figure 16. Figure 16: Reconstruction results for Spatial Functa and GFM on the convection equation with [PITH_FULL_IMAGE:figures/full_fig_p023_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 12 canonical work pages

  1. [38]

    Improved implicit neural representation with fourier reparameterized training

    Kexuan Shi, Xingyu Zhou, and Shuhang Gu. Improved implicit neural representation with fourier reparameterized training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25985–25994, 2024

    2024

  2. [1]

    SIAM, 2007

    Randall J LeVeque.Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems. SIAM, 2007

    2007

  3. [2]

    Cambridge university press, 2002

    Randall J LeVeque.Finite volume methods for hyperbolic problems, volume 31. Cambridge university press, 2002

    work page 2002

  4. [3]

    An introduction to the finite element method.New York, 27(14), 1993

    Junuthula Narasimha Reddy. An introduction to the finite element method.New York, 27(14), 1993

    work page 1993

  5. [4]

    Workshop report on basic research needs for scientific machine learning: Core technologies for artificial intelligence

    Nathan Baker, Frank Alexander, Timo Bremer, Aric Hagberg, Yannis Kevrekidis, Habib Najm, Manish Parashar, Abani Patra, James Sethian, Stefan Wild, et al. Workshop report on basic research needs for scientific machine learning: Core technologies for artificial intelligence. Technical report, USDOE Office of Science (SC), Washington, DC (United States), 2019

    2019

  6. [5]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    2021

  7. [6]

    Neural operators for accelerating scientific simulations and design

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandkumar. Neural operators for accelerating scientific simulations and design. Nature Reviews Physics, 6(5):320–328, 2024

    2024

  8. [7]

    Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92(3):88, 2022

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next.Journal of Scientific Computing, 92(3):88, 2022

    2022

Show all 44 references
  1. [8]

    Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    2023

  2. [9]

    Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    2022

  3. [10]

    Openfwi: Large-scale multi-structural benchmark datasets for full waveform inversion.Advances in Neural Information Processing Systems, 35:6007–6020, 2022

    Chengyuan Deng, Shihang Feng, Hanchen Wang, Xitong Zhang, Peng Jin, Yinan Feng, Qili Zeng, Yinpeng Chen, and Youzuo Lin. Openfwi: Large-scale multi-structural benchmark datasets for full waveform inversion.Advances in Neural Information Processing Systems, 35:6007–6020, 2022

    2022

  4. [11]

    Understanding training and generalization in deep learning by fourier analysis

    Zhiqin John Xu. Understanding training and generalization in deep learning by fourier analysis. arXiv preprint arXiv:1808.04295, 2018

    2018 arXiv

  5. [12]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. InInternational conference on machine learning, pages 5301–5310. PMLR, 2019

  6. [13]

    Overview frequency principle/spectral bias in deep learning.Communications on Applied Mathematics and Computation, pages 1–38, 2024

    Zhi-Qin John Xu, Yaoyu Zhang, and Tao Luo. Overview frequency principle/spectral bias in deep learning.Communications on Applied Mathematics and Computation, pages 1–38, 2024

    2024

  7. [14]

    Im- plicit neural representations with periodic activation functions.Advances in neural information processing systems, 33:7462–7473, 2020

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Im- plicit neural representations with periodic activation functions.Advances in neural information processing systems, 33:7462–7473, 2020

    2020

  8. [15]

    Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains.Advances in neural information proce...

    2020

  9. [16]

    Wire: Wavelet implicit neural representations

    Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507–18516, 2023. 10

    2023

  10. [17]

    From data to functa: Your data point is a function and you can treat it like one.arXiv preprint arXiv:2201.12204, 2022

    Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one.arXiv preprint arXiv:2201.12204, 2022

    2022 arXiv

  11. [18]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  12. [19]

    Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, and George Em Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193, 2019

    1910 arXiv

  13. [20]

    Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  14. [21]

    Factorized fourier neural operators

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized fourier neural operators. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  15. [22]

    Continuous pde dynamics forecasting with implicit neural representations.arXiv preprint arXiv:2209.14855, 2022

    Yuan Yin, Matthieu Kirchmeyer, Jean-Yves Franceschi, Alain Rakotomamonjy, and Patrick Gallinari. Continuous pde dynamics forecasting with implicit neural representations.arXiv preprint arXiv:2209.14855, 2022

    2022 arXiv

  16. [23]

    Gridmix: Exploring spatial modulation for neural fields in pde modeling

    Honghui Wang, Shiji Song, and Gao Huang. Gridmix: Exploring spatial modulation for neural fields in pde modeling. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  17. [24]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

    2019

  18. [25]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoor- thi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  19. [26]

    Neural tangent kernel: Convergence and generalization in neural networks.Advances in neural information processing systems, 31, 2018

    Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks.Advances in neural information processing systems, 31, 2018

    2018

  20. [27]

    Neural networks fail to learn periodic functions and how to fix it.Advances in Neural Information Processing Systems, 33:1583–1594, 2020

    Liu Ziyin, Tilman Hartwig, and Masahito Ueda. Neural networks fail to learn periodic functions and how to fix it.Advances in Neural Information Processing Systems, 33:1583–1594, 2020

    2020

  21. [28]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019

    2019

  22. [29]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021

    2021

  23. [30]

    Multiwavelet-based operator learning for differential equations.Advances in neural information processing systems, 34:24048–24062, 2021

    Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations.Advances in neural information processing systems, 34:24048–24062, 2021

    2021

  24. [31]

    Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

    Tapas Tripura and Souvik Chakraborty. Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

    2022 arXiv

  25. [32]

    Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020

    2003 arXiv

  26. [33]

    Learning mesh-based simulation with graph networks.arXiv preprint arXiv:2010.03409, 2020

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W Battaglia. Learning mesh-based simulation with graph networks.arXiv preprint arXiv:2010.03409, 2020. 11

    2010 arXiv

  27. [34]

    Fourier neural operator with learned deformations for pdes on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries.Journal of Machine Learning Research, 24(388):1–26, 2023

    2023

  28. [35]

    Geometry-informed neural operator for large-scale 3d pdes.Advances in Neural Information Processing Systems, 36, 2024

    Zongyi Li, Nikola Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Otta, Moham- mad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, et al. Geometry-informed neural operator for large-scale 3d pdes.Advances in Neural Information Processing Systems...

    2024

  29. [36]

    Operator learning with neural fields: Tackling pdes on general geometries.Advances in Neural Information Processing Systems, 36:70581–70611, 2023

    Louis Serrano, Lise Le Boudec, Armand Kassaï Koupaï, Thomas X Wang, Yuan Yin, Jean-Noël Vittaut, and Patrick Gallinari. Operator learning with neural fields: Tackling pdes on general geometries.Advances in Neural Information Processing Systems, 36:70581–70611, 2023

    2023

  30. [37]

    Spatial functa: Scaling functa to imagenet classification and generation

    Matthias Bauer, Emilien Dupont, Andy Brock, Dan Rosenbaum, Jonathan Richard Schwarz, and Hyunjik Kim. Spatial functa: Scaling functa to imagenet classification and generation. arXiv preprint arXiv:2302.03130, 2023

    2023 arXiv

  31. [39]

    An overview of full-waveform inversion in exploration geophysics.Geophysics, 74(6):WCC1–WCC26, 2009

    Jean Virieux and Stéphane Operto. An overview of full-waveform inversion in exploration geophysics.Geophysics, 74(6):WCC1–WCC26, 2009

    2009

  32. [40]

    Model-agnostic meta-learning for fast adap- tation of deep networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adap- tation of deep networks. InInternational conference on machine learning, pages 1126–1135. PMLR, 2017

    2017

  33. [41]

    Fast context adaptation via meta-learning

    Luisa M Zintgraf, Kyriacos Shiarlis, Vitaly Kurin, Katja Hofmann, and Shimon Whiteson. Fast context adaptation via meta-learning. 2019

    2019

  34. [42]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part I...

    2015

  35. [43]

    On the benefits of memory for modeling time-dependent PDEs

    Ricardo Buitrago, Tanya Marwah, Albert Gu, and Andrej Risteski. On the benefits of memory for modeling time-dependent PDEs. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  36. [44]

    Inducing point operator transformer: A flexible and scalable architecture for solving pdes

    Seungjun Lee and Taeil Oh. Inducing point operator transformer: A flexible and scalable architecture for solving pdes. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 153–161, 2024. 12 A Symbol Definitions Table 5 summarizes the key symbols an...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.