Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Feeding a fixed Chebyshev polynomial dictionary to the trunk, instead of raw coordinates, lets DeepONet-style operator networks resolve boundary layers and steep gradients better than coordinate-input or Fourier-feature trunks on bounded no

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Replacing the coordinate-input trunk of a DeepONet with a fixed Chebyshev polynomial dictionary lowers reported reconstruction error on bounded non-periodic PDE benchmarks.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection SEDONet is a plausible, cleanly motivated Chebyshev-trunk variant of DeepONet, but the empirical comparison is under-controlled and the 'parameter-neutral' claim is likely wrong as stated. the 3 major comments →

arxiv 2512.09165 v2 pith:FEEUZHCX submitted 2025-12-09 cs.LG

Spectral Embedding via Chebyshev Bases for Robust DeepONet Approximation

Muhammad Abid , Omer San This is my paper
classification cs.LG MSC 68T0765M70
keywords Chebyshev polynomialsDeep operator networksspectral embeddingsoperator learningPDE surrogatesnon-periodic domainsFourier featuresspectral bias
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Standard DeepONet trunks map raw coordinates through an MLP, a choice that inherits neural networks' spectral bias toward low frequencies and mismatches PDE solutions on bounded non-periodic domains. SEDONet replaces those coordinates with a fixed tensor-product Chebyshev polynomial dictionary before the trunk MLP, injecting a non-periodic spectral prior with no trainable parameters added to the embedding. On five benchmarks — Poisson, Burgers, advection-diffusion, Lorenz-96, and Allen-Cahn — the paper reports the lowest relative L2 errors among the three architectures compared, with average gains of 30-40% over the coordinate baseline and consistent gains over a Fourier-embedded variant. The point is that choosing the spectral family that matches the domain's boundary structure matters as much as network capacity.

Core claim

The central claim is that a fixed Chebyshev dictionary, not learned, entered as the trunk input, changes what a DeepONet can efficiently represent on bounded domains. The trunk then outputs learned nonlinear mixtures of Chebyshev modes, which the branch contracts against input-dependent coefficients, mirroring a classical spectral expansion. Across all five benchmarks the reported relative L2 errors are lowest for SEDONet (Poisson 0.97% vs 1.39% for DeepONet; Burgers 3.98% vs 6.20%; advection-diffusion 4.33% vs 8.00%; Lorenz-96 20.90% vs 23.63%; Allen-Cahn 5.44% vs 10.01%), and energy spectrum analyses show the Chebyshev trunk preserves intermediate- and high-frequency content better than th

What carries the argument

The key object is the fixed tensor-product Chebyshev dictionary Φ_ij(x,t)=T_i(2x−1)T_j(2t−1), where T_n are Chebyshev polynomials of the first kind defined by T_0=1, T_1=ξ, T_{n+1}=2ξT_n−T_{n−1}. Its entries are stacked, cropped or padded to the trunk's input dimension, and fed to the trunk MLP instead of raw coordinates; nothing in the embedding is learned. The dictionary supplies near-orthogonal, well-conditioned features on [−1,1] with node clustering at the boundaries, giving the trunk direct access to high-degree polynomial modes that coordinate-input MLPs learn inefficiently. This embedding carries the argument by making the spectral prior explicit and parameter-free.

Load-bearing premise

The load-bearing premise is that the benchmark comparison isolates the Chebyshev prior: because dictionary sizes, trunk input dimension, layer widths, and training hyperparameters are not reported, the gains could partly reflect a larger or differently tuned trunk rather than the spectral embedding itself.

What would settle it

Run the identical SEDONet pipeline with the Chebyshev dictionary replaced by a fixed random-feature dictionary of the same input dimension and the same trainable parameter count; if random features reproduce the reported errors, the Chebyshev structure is not doing the work. Reporting K_x, K_t, d_trunk, layer widths, and matching total parameter counts would make this test decisive.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Bounded-domain operator learning gains a natural default embedding: choose Chebyshev rather than Fourier when the PDE boundary conditions are non-periodic.
  • Because the embedding is deterministic and has no learnable parameters, the modification is a drop-in replacement for any DeepONet trunk and can be combined with other branch or trunk improvements.
  • Energy spectra of SEDONet predictions stay closer to the reference in intermediate and high frequencies, so downstream quantities that depend on gradients and fine scales should also improve.
  • The tensor-product construction extends to higher spatial dimensions and to space-time inputs, so the same spectral-prior argument carries over to 3-D and time-dependent problems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable corollary the paper leaves implicit: other fixed orthogonal polynomial families matched to domain boundary conditions, such as Legendre polynomials, should produce similar gains — which would show the benefit is non-periodic spectral whitening rather than Chebyshev per se.
  • The 'parameter-neutral' label is narrower than it sounds: replacing raw coordinates with a K_x by K_t dictionary changes the first trunk layer's input dimension and parameter count, so a fully controlled comparison needs matching parameter budgets and reported hyperparameters.
  • If the spectral-bias interpretation is right, the margin over coordinate trunks should widen as solutions become stiffer or more boundary-dominated, suggesting a diagnostic sweep the current benchmarks only partially cover.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces SEDONet, a DeepONet variant in which the trunk network receives a fixed, tensor-product Chebyshev polynomial embedding of the spatio-temporal coordinates instead of raw coordinate inputs. The embedding is deterministic and adds no trainable parameters of its own. The authors argue that this provides a non-periodic spectral inductive bias suited to bounded domains, in contrast to Fourier features. They compare SEDONet with standard DeepONet and FEDONet on five benchmarks: 2-D Poisson, 1-D Burgers, 1-D advection-diffusion, Lorenz-96, and Allen-Cahn. The paper reports lower mean relative L2 errors for SEDONet on all benchmarks, presents spectral energy comparisons, and includes an appendix claiming that the Chebyshev-embedded hypothesis class strictly contains the vanilla coordinate-input class.

Significance. If the empirical claims are supported, SEDONet would be a simple, practical contribution: a drop-in spectral embedding for DeepONet trunks that improves accuracy on bounded, non-periodic problems while preserving the branch-trunk factorization. The idea of using a fixed Chebyshev dictionary is well motivated by classical spectral methods and is easy to implement in existing DeepONet codes. However, the current manuscript does not yet establish the central empirical claim because the experimental comparison is not controlled: the Chebyshev embedding changes the input dimension of the first trunk layer, and the paper does not report the embedding sizes, network widths, or training hyperparameters. In addition, several of the reported gains are within one standard deviation of the FEDONet baseline and no significance testing is provided. The theoretical appendix does not provide a rigorous proof of the claimed strict hypothesis-class enlargement. These issues are fixable, but they must be addressed before the paper's conclusions can be accepted.

major comments (3)
  1. [§2.3, Eq. (10); §2.4; Table 1] The 'parameter-neutral' claim is repeated throughout the paper, but the construction in Eq. (10) replaces the raw coordinate input with a d_trunk-dimensional Chebyshev feature vector. Unless d_trunk equals the coordinate-space dimension, the first trunk layer has a different number of trainable parameters. The paper never reports K_x, K_t, d_trunk, the crop/pad rule C, the trunk/branch widths and depths, or the training hyperparameters. The observed gains in Table 1 could therefore reflect a larger/richer trunk rather than the Chebyshev prior specifically. The authors should report the full architecture and hyperparameter configuration, give total parameter counts for each model, and include capacity-matched baselines (e.g., vanilla DeepONet and FEDONet with comparable parameter counts or comparable first-layer widths). Without this, the central claim that the Chebyshev embedding is the
  2. [Table 1; §3.2; §3.5] The abstract and Section 4 describe the results as 'lowest or statistically comparable,' but no significance tests or multi-seed repetitions are reported. Several margins are small relative to the reported standard deviations: Burgers SEDONet 3.98±2.64 vs. FEDONet 4.47±3.08, and Allen-Cahn SEDONet 5.44±3.05 vs. FEDONet 5.70±3.57. The paper should report results over multiple independent training runs and either paired significance tests or confidence intervals. As written, the claim of consistent improvement over FEDONet on these benchmarks is not statistically supported.
  3. [Appendix B; §2.3] The paper claims that the Chebyshev embedding 'strictly enlarges' the hypothesis space, but Appendix B is explicitly an informal argument rather than a proof. The inclusion step uses polynomial approximation and a 'single linear neuron' on top of the Chebyshev features, which does not establish inclusion for a fixed-width/depth coordinate network. The strictness argument similarly lacks a concrete capacity measure or width/depth bounds. Since this claim is used to justify the design, the authors should either provide a rigorous statement with clear architectural assumptions or soften the claim to say that the Chebyshev embedding provides a more direct representation of high-order polynomial modes. The current text overstates what Appendix B demonstrates.
minor comments (6)
  1. [Abstract; Table 1] The abstract and Section 1 mention improvements 'up to 54%' over baseline DeepONet, but the largest relative reduction in Table 1 is about 46% (advection-diffusion, 8.00% to 4.33%). The 54% figure does not appear to be supported by the reported results and should be reconciled or removed.
  2. [§3.5] The Allen-Cahn benchmark uses periodic boundary conditions on [-1,1]. The paper's main narrative is about non-periodic bounded domains, so the small SEDONet gain on this periodic problem is not explained by the proposed prior. The authors should either justify why Chebyshev features help here or explicitly frame Allen-Cahn as a stress test outside the intended regime.
  3. [Eq. (10)] The crop/pad operator C is not specified. It is important for reproducibility to state which modes are cropped or padded and how d_trunk is chosen relative to K_x K_t.
  4. [Eq. (16)] The notation G_theta(u) and G(u) in Eq. (16) is inconsistent with the earlier use of G_theta and G: the mathematical operator is G and the predictor is G_theta. Please clarify the subscript conventions.
  5. [Appendix A] The claim that the Gram matrix is 'close to diagonal and well-conditioned' when coordinates are near Gauss-Lobatto points needs a precise statement or a small numerical demonstration. As written, it is a qualitative assertion.
  6. [Throughout] There are numerous typographical and grammatical errors, e.g., 'a follows' before Eq. (8), 'as the followings', and the use of 'parameter-neutral' after introducing a parameter-free embedding without discussing input dimension. A careful language edit is needed.

Circularity Check

0 steps flagged

No significant circularity: benchmark claims are empirically self-contained; remaining caveats are reporting and proof-strength issues, not circular reductions.

full rationale

The paper's derivation chain is: define a fixed tensor-product Chebyshev dictionary (Eqs. 5-10), feed it to the trunk MLP, keep the branch unchanged, synthesize via the DeepONet inner product (Eq. 13), and train by MSE (Eq. 14). The Chebyshev embedding phi_Cheb is deterministic and has no trainable parameters, so the relative L2 errors in Table 1 are external experimental outputs rather than recovered inputs. No parameter is fitted to the test set and then renamed a prediction; no uniqueness theorem is imported from the authors' prior work; no ansatz is adopted solely through a self-citation. The FEDONet reference [59] (which shares an author) is used only as a baseline architecture, not as load-bearing evidence for the Chebyshev claim. Appendix B's 'superset' argument is admittedly 'intentionally high level and is meant to provide intuition rather than a fully measure-theoretic proof,' and its exact representability of T_K is by construction of the feature vector; this is close to tautological and does not by itself establish strict set inclusion, but it is not used to generate or tune the empirical benchmark results. The 'parameter-neutral' claim in Section 2.3 and the unreported K_x, K_t, d_trunk, and widths leave a potential capacity confound; that is an experimental reporting/control issue, not a circular derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced; the Chebyshev embedding is a fixed deterministic feature map. The central empirical claim rests mainly on undisclosed hyperparameters, the fairness of the training protocol, and the accuracy of solver-generated reference data.

free parameters (4)
  • Chebyshev orders K_x, K_t and trunk embedding dimension d_trunk
    Chosen by hand and not reported; they determine the trunk feature dimension in Eqs. (8)-(10) and are central to the parameter-neutrality claim.
  • Crop/pad operator C
    Eq. (10): C truncates or zero-pads the Chebyshev dictionary to match the trunk input size; the choice affects which spectral modes are retained.
  • Trunk/branch MLP widths, depth, and trunk channels p
    Not reported; these define model capacity and affect the fairness of comparisons across DeepONet, FEDONet, and SEDONet.
  • Training hyperparameters (learning rate, epochs, batch size, seeds)
    Not reported; no multi-seed variance is given, so the statistical significance of the reported differences cannot be assessed.
axioms (5)
  • standard math Chebyshev polynomials are orthogonal on [-1,1] with respect to the weight (1-xi^2)^(-1/2).
    Appendix A uses this classical orthogonality to argue the feature Gram matrix is well conditioned.
  • standard math Coordinate-input MLP trunks are universal approximators but exhibit spectral bias.
    Section 2.2 relies on universal approximation and spectral-bias results from the cited literature to motivate the Chebyshev embedding.
  • domain assumption Sampling coordinates at or near Gauss-Lobatto points makes the empirical Chebyshev feature matrix nearly orthogonal.
    Appendix A asserts near-orthogonality for points close to Gauss-Lobatto collocation; the actual test grids are not shown to satisfy this.
  • domain assumption DeepONet, FEDONet, and SEDONet are trained under a matched protocol with comparable capacity and training budget.
    Section 2.4 states a common protocol, but no hyperparameters, parameter counts, or seeds are reported to support the assumption.
  • domain assumption The finite-difference and RK4 reference solutions are accurate enough to serve as ground truth for relative L2 error evaluation.
    All benchmarks rely on numerically generated reference data; solver errors are not quantified or compared with model errors.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral Embedding via Chebyshev Bases for Robust DeepONet Approximation." pith.science (2026). https://pith.science/paper/FEEUZHCX

@misc{pith2026251209165,
  author       = {Pith},
  title        = {Pith review of: Spectral Embedding via Chebyshev Bases for Robust DeepONet Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FEEUZHCX}},
  note         = {Machine review of arXiv:2512.09165}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep Operator Networks (DeepONets) have emerged as a powerful framework for data-driven operator learning, providing flexible surrogates for nonlinear mappings arising in partial differential equations (PDEs). However, the standard trunk network, which operates directly on raw spatial or spatiotemporal coordinates through fully connected layers, often struggles to represent sharp gradients, boundary layers, and other non-periodic solution structures on bounded domains. To address these limitations, we introduce the Spectral-Embedded Deep Operator Network (SEDONet), a novel DeepONet architecture in which the trunk is driven by a fixed Chebyshev spectral dictionary instead of coordinate inputs. This non-periodic spectral embedding provides a principled inductive bias for bounded domains, enabling the learned operator to capture fine-scale features that are difficult for Fourier-based or MLP-only trunks to represent. SEDONet is evaluated on the 2-D Poisson equation, 1-D Burgers' equation, 1-D advection-diffusion equation, Allen-Cahn equation, Lorenz-96 chaotic system, and Darcy flow, covering elliptic, hyperbolic, parabolic, chaotic, and multiscale problems. Across all benchmarks, SEDONet consistently achieves the lowest or statistically comparable relative $L^2$ errors among DeepONet, FEDONet, and SEDONet, with improvements of up to 54% over the baseline DeepONet and consistent gains over Fourier-embedded variants on bounded, non-periodic problems. Energy spectrum analyses further demonstrate that SEDONet more accurately preserves intermediate- and high-frequency solution structures. The proposed framework provides a simple, parameter-neutral modification to DeepONets, offering a robust and computationally efficient spectral approach for surrogate modeling of nonlinear operators in scientific computing.

Figures

Figures reproduced from arXiv: 2512.09165 by Muhammad Abid, Omer San.

Figure 1
Figure 1. Figure 1: SEDONet architecture: The branch network maps discrete samples of the input function u0 to latent coefficients bk, while the trunk network maps Chebyshev spectral features Φ(x, t) to basis channels tk(x, t). Their inner product yields the operator evaluation Gb(u0)(x, t). 6 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of DeepONet, FEDONet, and SEDONet on a representative test example from the 2D Poisson [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relative ℓ2 error (mean ± std) across 1000 unseen Poisson test set for all three architectures. SEDONet achieves both the lowest mean error and the lowest variance. These qualitative observations are consistent with the quantitative results. Across the full test set, DeepONet attains a mean relative ℓ2 error of 1.39%, FEDONet reduces this to 1.10%, and SEDONet further improves the accuracy to 0.97%. This c… view at source ↗
Figure 4
Figure 4. Figure 4: Relative ℓ2 error across 128 unseen Burgers’ test samples for all three architectures (DeepONet, FEDONet, and SEDONet). We can see that a test example is shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Best-performing Burgers’ test sample: Ground truth, DeepONet, FEDONet, and SEDONet predictions with [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Energy spectrum comparison for Burgers’ equation: temporal evolution of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative test sample from the non-periodic advection-diffusion dataset. DeepONet exhibits smoothing [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relative ℓ2 error (mean ± std) across 250 unseen advection-diffusion test samples. SEDONet achieves the lowest error and improved robustness compared to both DeepONet and FEDONet. Overall, this benchmark highlights that SEDONet provides a clear advantage when learning operators associated with non-periodic, boundary-constrained PDEs. The Chebyshev-based trunk offers a more suitable spectral representation … view at source ↗
Figure 9
Figure 9. Figure 9: Spatio-temporal reconstruction of representative Lorenz-96 trajectories: ground truth, DeepONet, FEDONet, [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Relative ℓ2 error (mean ± std) across 2000 unseen Lorenz-96 trajectories. SEDONet achieves both the lowest average error and the lowest variance, demonstrating improved stability and robustness compared to DeepONet and FEDONet. Overall, these results indicate that even for smooth, quasi-periodic systems like Lorenz-96, where fully connected trunks already perform reasonably well, the Chebyshev spectral em… view at source ↗
Figure 11
Figure 11. Figure 11: Spatio-temporal reconstruction for a representative Allen-Cahn test sample: ground truth, DeepONet, [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Relative ℓ2 error (mean ± std) across 250 unseen Allen–Cahn test samples for DeepONet, FEDONet, and SEDONet. SEDONet achieves the lowest error with reduced variance across the ensemble. These qualitative observations are confirmed quantitatively by the results summarized in [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. WLNO: Wavelet-Laplace Neural Operator for Solving Partial Differential Equations

    cs.LG 2026-05 unverdicted novelty 6.0

    WLNO augments LNO with a parallel Haar wavelet branch and learnable gate to capture multi-scale spatial features, outperforming LNO on five PDE benchmarks especially those with sharp structures.

  2. The impact of observation density on Bayesian inversion of latent dynamics in shock-dominated flows

    cs.LG 2026-05 unverdicted novelty 5.0

    A latent-space reduced-order model using autoencoders and learned dynamics enables Bayesian recovery of initial density and pressure in Sod shock tube simulations, with posterior uncertainty contracting substantially ...

Reference graph

Works this paper leans on

66 extracted references · 17 linked inside Pith · cited by 2 Pith papers

  1. [1]

    J. P. Boyd, Chebyshev and Fourier Spectral Methods, Courier Corporation, 2001

    2001

  2. [2]

    Gottlieb, S

    D. Gottlieb, S. A. Orszag, Numerical Analysis of Spectral Methods: Theory and Applications, SIAM, 1977. doi:10.1137/1.9781611970425

    doi:10.1137/1.9781611970425 1977

  3. [3]

    W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Numerical Recipes: The Art of Scientific Computing, Cambridge University Press, 1986

    1986

  4. [4]

    Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals and Systems 2 (4) (1989) 303–314

    G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals and Systems 2 (4) (1989) 303–314

    1989

  5. [5]

    Hornik, M

    K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural Networks 2 (5) (1989) 359–366

    1989

  6. [6]

    Poggio, F

    T. Poggio, F. Girosi, Networks for approximation and learning, Proceedings of the IEEE 78 (9) (1990) 1481–1497

    1990

  7. [7]

    T. Chen, H. Chen, Universal approximation to nonlinear operators by neural networks, IEEE Transactions on Neural Networks 6 (4) (1995) 911–917

    1995

  8. [8]

    Sirovich, Turbulence and the dynamics of coherent structures, Quarterly of Applied Mathematics 45 (3) (1987) 561–571

    L. Sirovich, Turbulence and the dynamics of coherent structures, Quarterly of Applied Mathematics 45 (3) (1987) 561–571

    1987

  9. [9]

    Berkooz, P

    G. Berkooz, P. Holmes, J. L. Lumley, The proper orthogonal decomposition in the analysis of turbulent flows, Annual Review of Fluid Mechanics 25 (1993) 539–575

    1993

  10. [10]

    C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for Machine Learning, MIT Press, 2005. 16 APREPRINT

    2005

  11. [11]

    E. J. Kansa, Multiquadrics for scattered data approximation, Computers & Mathematics with Applications 19 (8) (1990) 127–145

    1990

  12. [12]

    D. Lowe, D. Broomhead, Multivariable functional interpolation and adaptive networks, Complex Systems 2 (3) (1988) 321–355

    1988

  13. [13]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet, Nature Machine Intelligence 3 (2021) 218–229

    2021

  14. [14]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces, Journal of Machine Learning Research 24 (1) (2023) 1–97

    2023

  15. [15]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric PDEs (2021).arXiv:2010.08895

    Pith/arXiv arXiv 2021

  16. [16]

    Tripura, S

    T. Tripura, S. Chakraborty, Wavelet neural operator, Computer Methods in Applied Mechanics and Engineering 404 (2023) 115783

    2023

  17. [17]

    Gupta, X

    G. Gupta, X. Xiao, P. Bogdan, Multiwavelet-based operator learning, in: Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  18. [18]

    Kossaifi, N

    J. Kossaifi, N. Kovachki, K. Azizzadenesheli, A. Anandkumar, Multi-grid tensorized fourier neural operator, arXiv preprint arXiv:2310.00120 (2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Guo, H.-B

    Z.-H. Guo, H.-B. Li, MgFNO: Multi-grid architecture fourier neural operator (2024).arXiv:2407.08615

    Pith/arXiv arXiv 2024

  20. [20]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Graph kernel network (2020).arXiv:2003.03485

    Pith/arXiv arXiv 2020

  21. [21]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Multipole graph neural operator (2020).arXiv:2006.09535

    Pith/arXiv arXiv 2020

  22. [22]

    Z. Li, D. Z. Huang, B. Liu, A. Anandkumar, Fourier neural operator with learned deformations, Journal of Machine Learning Research 24 (388) (2023) 1–26

    2023

  23. [23]

    Z. Li, N. Kovachki, C. Choy, B. Li, J. Kossaifi, S. Otta, M. A. Nabian, M. Stadler, C. Hundt, K. Azizzadenesheli, et al., Geometry-informed neural operator, Advances in Neural Information Processing Systems (NeurIPS) (2023)

    2023

  24. [24]

    Huang, K

    J. Huang, K. Zhang, Y . Wu, Z. Cheng, Operator learning with domain decomposition for geometry generalization (2025).arXiv:2504.00510

    arXiv 2025

  25. [25]

    S. E. Ahmed, P. Stinis, A multifidelity deep operator network approach to closure for multiscale systems, Computer Methods in Applied Mechanics and Engineering 414 (2023) 116161.doi:10.1016/j.cma.2023.116161

    arXiv 2023

  26. [26]

    A. A. Howard, M. Perego, G. E. Karniadakis, P. Stinis, Multifidelity deep operator networks for data-driven and physics-informed problems, Journal of Computational Physics 493 (2023) 112462. doi:10.1016/j.jcp.2023. 112462

    doi:10.1016/j.jcp.2023 2023

  27. [27]

    Batlle, M

    P. Batlle, M. Darcy, B. Hosseini, H. Owhadi, Kernel methods are competitive for operator learning, Journal of Computational Physics 496 (2024) 112549.doi:10.1016/j.jcp.2023.112549

    arXiv 2024

  28. [28]

    Kumar, R

    S. Kumar, R. Nayek, S. Chakraborty, Neural operator induced gaussian process framework for probabilistic solution of parametric partial differential equations, Computer Methods in Applied Mechanics and Engineering 431 (2024) 117265.doi:10.1016/j.cma.2024.117265

    arXiv 2024

  29. [29]

    S. Garg, S. Chakraborty, Neuroscience inspired neural operator for partial differential equations, Journal of Computational Physics 515 (2024) 113266.doi:10.1016/j.jcp.2024.113266

    arXiv 2024

  30. [30]

    O’Leary-Roseberry, Y

    T. O’Leary-Roseberry, Y . Chen, U. Villa, O. Ghattas, Derivative-informed neural operator: An interpretable neural operator architecture for learning parametric differential operators, Journal of Computational Physics 496 (2024) 112555.doi:10.1016/j.jcp.2023.112555

    arXiv 2024

  31. [31]

    X. Liu, B. Xu, S. Cao, L. Zhang, Mitigating spectral bias for the multiscale operator learning, Journal of Computational Physics 506 (2024) 112944.doi:10.1016/j.jcp.2024.112944

    arXiv 2024

  32. [32]

    S. Wu, A. Zhu, Y . Tang, B. Lu, Solving parametric elliptic interface problems via interfaced operator network (IONet), Journal of Computational Physics 514 (2024) 113217.doi:10.1016/j.jcp.2024.113217

    arXiv 2024

  33. [33]

    X. Bi, X. Chen, C. Zhao, Q. Li, J. Zhang, XI-DeepONet: An operator learning method for elliptic interface problems, Journal of Computational Physics 538 (2025) 114164.doi:10.1016/j.jcp.2024.114164

    arXiv 2025

  34. [35]

    Huang, Y

    J. Huang, Y . Qiu, Resolution invariant deep operator network for PDEs with complex geometries, Journal of Computational Physics 522 (2025) 113601.doi:10.1016/j.jcp.2024.113601

    arXiv 2025

  35. [36]

    Y . Meng, J. Huang, Y . Qiu, Koopman operator learning using invertible neural networks, Journal of Computational Physics 501 (2024) 112795.doi:10.1016/j.jcp.2024.112795

    arXiv 2024

  36. [38]

    W. Chen, P. Stinis, Feature-adjacent multi-fidelity physics-informed machine learning for partial differential equations, Journal of Computational Physics 498 (2024) 112683.doi:10.1016/j.jcp.2023.112683

    arXiv 2024

  37. [39]

    Jafarzadeh, S

    S. Jafarzadeh, S. Silling, N. Liu, Z. Zhang, Y . Yu, Peridynamic neural operators: A data-driven nonlocal constitutive model for complex material responses, Computer Methods in Applied Mechanics and Engineering 425 (2024) 116914.doi:10.1016/j.cma.2024.116914

    arXiv 2024

  38. [40]

    X. Dong, C. Chen, J.-L. Wu, Data-driven stochastic closure modeling via conditional diffusion model and neural operator, Journal of Computational Physics 534 (2025) 114005.doi:10.1016/j.jcp.2025.114005

    arXiv 2025

  39. [41]

    Bahmani, S

    B. Bahmani, S. Goswami, I. G. Kevrekidis, M. D. Shields, A resolution independent neural operator, Computer Methods in Applied Mechanics and Engineering 444 (2025) 118113

    2025

  40. [42]

    H. Li, X. Ye, P. Jiang, G. Qin, T. Wang, Local neural operator for solving transient partial differential equations on varied domains, Computer Methods in Applied Mechanics and Engineering 427 (2024) 117062. doi: 10.1016/j.cma.2024.117062

    arXiv 2024

  41. [43]

    Zhong, H

    W. Zhong, H. Meidani, Physics-informed geometry-aware neural operator, Computer Methods in Applied Mechanics and Engineering 415 (2023) 116278.doi:10.1016/j.cma.2023.116278

    arXiv 2023

  42. [44]

    X. Xu, J. Guilleminot, V . Tarokh, Neural operators from the cole-hopf transformation: Leveraging relations between pdes for efficient operator learning, Computer Methods in Applied Mechanics and Engineering (2025)

    2025

  43. [45]

    Goswami, Z

    S. Goswami, Z. Li, K. Azizzadenesheli, M. Liu-Schiaffini, K. Bhattacharya, H. Hassani, A. M. Stuart, A. Anand- kumar, Physics-informed deep operator networks (2022).arXiv:2207.05748

    Pith/arXiv arXiv 2022

  44. [46]

    Z. Li, N. Kovachki, K. Azizzadenesheli, K. Bhattacharya, A. M. Stuart, A. Anandkumar, Physics-informed neural operator (2023).arXiv:2309.15502

    Pith/arXiv arXiv 2023

  45. [47]

    M. S. Eshaghi, P. Akhtari, G. E. Karniadakis, Variational physics-informed neural operator (2024). arXiv: 2403.08377

    Pith/arXiv arXiv 2024

  46. [48]

    K. Chen, M. Zhao, W. Li, G. Lin, Pseudo physics-informed neural operators (2025).arXiv:2501.01234

    Pith/arXiv arXiv 2025

  47. [49]

    T. Wang, C. Wang, Latent neural operators (2024).arXiv:2404.00967

    Pith/arXiv arXiv 2024

  48. [50]

    T. Wang, C. Wang, Latent neural operator pretraining (2024).arXiv:2408.05677

    Pith/arXiv arXiv 2024

  49. [51]

    Ahmad, N

    Z. Ahmad, N. Kovachki, Z. Li, A. M. Stuart, A. Anandkumar, Diffeomorphic latent neural operators (2024). arXiv:2406.01411

    Pith/arXiv arXiv 2024

  50. [52]

    D. Long, Y . Rao, J. Lu, J. Zhou, Invertible fourier neural operators (2025).arXiv:2502.04521

    arXiv 2025

  51. [53]

    N. Cho, J. Ryu, H. J. Hwang, Sobolev training for operator learning, Journal of Computational Physics 543 (2025) 114408.doi:10.1016/j.jcp.2025.114408

    arXiv 2025

  52. [54]

    Huang, R

    H. Huang, R. Lai, Unsupervised solution operator learning for mean-field games, Journal of Computational Physics 537 (2025) 114057.doi:10.1016/j.jcp.2025.114057

    arXiv 2025

  53. [55]

    L. Yang, S. Osher, Pde generalization of in-context operator networks: A study on 1d scalar nonlinear conservation laws, Journal of Computational Physics 519 (2024) 113379.doi:10.1016/j.jcp.2024.113379

    arXiv 2024

  54. [56]

    Rahaman, A

    N. Rahaman, A. Baratin, D. Arpit, F. Draxler, M. Lin, F. A. Hamprecht, Y . Bengio, A. Courville, On the spectral bias of neural networks (2019).arXiv:1806.08734

    Pith/arXiv arXiv 2019

  55. [57]

    Tancik, P

    M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, R. Ng, Fourier features for learning high frequency functions (2020).arXiv:2006.10739

    Pith/arXiv arXiv 2020

  56. [58]

    M. Kast, J. S. Hesthaven, Positional embeddings for solving PDEs, Journal of Computational Physics 505 (2024) 112924

    2024

  57. [59]

    Sojitra, M

    A. Sojitra, M. Dhingra, O. San, Fedonet: Fourier-embedded deeponet for spectrally accurate operator learning, arXiv preprint arXiv:2509.12344 (2025). 18 APREPRINT

    Pith/arXiv arXiv 2025

  58. [60]

    Z. Liu, H. Wang, H. Zhang, K. Bao, X. Qian, S. Song, Render unto numerics: Orthogonal polynomial neural operator for pdes with nonperiodic boundary conditions, SIAM Journal on Scientific Computing 46 (4) (2024) C323–C348

    2024

  59. [61]

    S. Mall, S. Chakraverty, Single layer Chebyshev neural network, Neural Computing and Applications 28 (2017) 915–929

    2017

  60. [62]

    S. M. Sivalingam, P. Kumar, V . Govindaraj, Chebyshev neural network for fractional PDEs, Computers & Mathematics with Applications (2024)

    2024

  61. [63]

    P. Yin, S. Ling, W. Ying, Chebyshev spectral neural networks for solving partial differential equations, arXiv preprint arXiv:2407.03347 (2024)

    Pith/arXiv arXiv 2024

  62. [64]

    Huang, H

    Y . Huang, H. Liu, Y . Zhao, M. Fei, Chebyshev spectral approximation-based physics-informed neural network for solving higher-order nonlinear differential equations, Engineering with Computers 41 (2) (2025) 1191–1210

    2025

  63. [65]

    S. Chen, S. Xiong, Y . Liu, Chebyshev-sobolev physics-informed neural networks for general pde solutions, International Journal of Applied and Computational Mathematics 11 (5) (2025) 169

    2025

  64. [66]

    Zhang, Y

    H. Zhang, Y . Wang, AC-PKAN: Attention-enhanced and chebyshev polynomial kolmogorov-arnold networks for physics-informed learning, in: Proceedings of the International Conference on Learning Representations (ICLR), 2024, openReview preprint

    2024

  65. [67]

    Z. Xu, Y . Chen, D. Xiu, Chebyshev feature neural network for accurate function approximation, arXiv preprint arXiv:2409.19135 (2024)

    Pith/arXiv arXiv 2024

  66. [68]

    near-orthogonality

    V . S. Fanaskov, I. V . Oseledets, Spectral neural operators, in: Doklady Mathematics, V ol. 108, Springer, 2023, pp. S226–S232. A Spectral Properties of Chebyshev Embeddings In this appendix we summarize a few classical spectral properties of Chebyshev polynomials that underpin the design of the SEDONet trunk. The goal is not to provide an exhaustive rev...

    2023

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.