Pith. sign in

REVIEW 4 major objections 4 minor 47 references

The paper claims that factorizing a spectral neural operator into an equivariant dynamic branch and an invariant persistent branch is a general computational principle for heterogeneous physical systems, yielding more accurate, parameter-ef

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 11:06 UTC pith:6XH2ZB7Y

load-bearing objection FaNO is a useful architectural trick—a cheap persistent branch that scales a fixed spectral template by the input's integral—but the paper oversells the 'spontaneous specialization' as a discovery when it is built into the equations. the 4 major comments →

arxiv 2606.16900 v2 pith:6XH2ZB7Y submitted 2026-06-15 cs.LG

Factorized Neural Operators Decompose Dynamic and Persistent Responses

Hao Tang , Yuechen Duan , Jiongyu Zhu , Zimeng Feng , Hao Li , Chao Li This is my paper
classification cs.LG
keywords neural operatorGreen's functionfactorized architectureequivarianceinvariancespectral methodsautoregressive forecastingcross-resolution generalization
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.

This paper tries to establish that heterogeneous physical systems—where fast transients coexist with long-lived structures—are better modeled by factorized neural operators than by single-branch spectral operators. It proposes FaNO, which splits the operator into an equivariant dynamic branch and an invariant persistent branch, and shows across spheres, grids, and manifolds that the two branches spontaneously specialize. If correct, this gives a general computational principle: explicit mechanism factorization improves accuracy, parameter efficiency, long-horizon stability, and cross-resolution generalization. The authors argue the benefits come from the factorization itself, not from added capacity.

Core claim

The central claim is that a physical system's response can be meaningfully decomposed into two operator components: g(u) = F⁻¹(concat[F[f]·G_E, C_f·G_I])·w₁, where G_E is an equivariant spectral kernel capturing transformation-sensitive transients and G_I is an invariant spectral kernel modulated only by the geometry-aware integral of the input, capturing persistent structure. The paper provides evidence that this decomposition is not merely architectural: across shallow-water dynamics, Navier–Stokes turbulence, cylinder wake flows, and other tasks, the two branches specialize spontaneously. The persistent branch acts as a stable structural anchor tied to geometry and boundary conditions, wh

What carries the argument

The load-bearing object is a pair of designed Green's functions. A relative-position kernel G(T⁻¹u) gives the dynamic branch g_E = F⁻¹[G_Eᶿ · F[f]], which is a convolution in the domain's native spectral basis (Fourier, spherical harmonics, or Laplace–Beltrami eigenfunctions). An absolute-position kernel G(u) gives the persistent branch g_I = F⁻¹[C_f · G_Iᶿ], where C_f is the geometry-aware integral (a scalar) of the input; because C_f is a scalar, this branch can only imprint a fixed spatial template scaled by global input intensity. The two branches are concatenated in spectral space and mixed by a linear layer, preserving the backbone architecture and channel budget of the single-branch o

Load-bearing premise

The persistent branch assumes that all persistent structure can be represented as one static spatial template scaled by the input's total integral; if persistent structure depends on where the input's mass sits, rather than only on how much total mass there is, the claimed decomposition cannot represent it.

What would settle it

Train FaNO on a synthetic system whose target is the sum of a transient wave and a persistent pattern whose spatial location or amplitude depends on the input's spatial distribution (for example, a Gaussian bump whose center is set by the input's first moment). If the persistent branch—forced to use only the scalar C_f—cannot represent the moving bump, and the full model's error grows with bump displacement while a single-branch operator does not, the claim that factorization is generally more accurate would be overturned.

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

If this is right

  • If FaNO's central claim holds, physical operator learning should not treat a system's response as one monolithic spectrum; decomposing by transformation behavior is a broadly applicable design choice.
  • The persistent branch's role as a structural anchor explains why autoregressive rollout stays stable: each step corrects transients around a slowly varying template instead of relearning global structure.
  • Because factorization itself, not capacity, drives the gains, the same backbone can be made smaller and cheaper while matching or beating single-branch accuracy.
  • The spontaneous specialization of branches into transient versus persistent responses gives an interpretable handle on what a learned operator believes about a physical system.
  • The framework extends beyond time-dependent PDEs to elliptic problems and geometric learning tasks, suggesting the dynamic/persistent split is a general representational principle.

Where Pith is reading between the lines

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

  • If the persistent branch is truly limited to C_f-scaled templates, then the framework's generality depends on persistent structure being representable as a single spatial pattern per channel; a natural test is to construct systems where two persistent patterns trade off according to input spatial distribution and see whether FaNO still captures them.
  • The paper's mechanism suggests a design recipe for other heterogeneity types: identify the transformation behavior of each response component and instantiate a corresponding Green's function branch, potentially adding scale-equivariant or approximately-equivariant branches for multiphysics.
  • The invariance-by-construction branch might especially help low-data regimes, since it encodes a strong prior that some structure is fixed; this could be tested by training FaNO and its single-branch baseline on small fractions of the training sets.
  • The learned persistent kernel G_Iᶿ can be read as an empirical 'system fingerprint': comparing G_Iᶿ across domains or viscosities might reveal whether the branch encodes boundary and geometry information or merely absorbs mean-field bias.

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

4 major / 4 minor

Summary. The paper proposes Factorized Neural Operators (FaNO), a spectral architecture that splits the learned operator into an equivariant-inspired dynamic branch S_E and an invariant-inspired persistent branch S_I, fused by channel mixing. The authors introduce a 'Unified Green's Function Framework' to motivate the two branch designs, derive equations for Euclidean, spherical, and manifold domains, and report experiments across SWE, WeatherBench, Navier–Stokes, CylinderFlow, Darcy, Helmholtz, Spherical MNIST, and surface segmentation. The central claims are that (i) the two branches spontaneously specialize into transient versus persistent responses, (ii) this factorization improves accuracy, parameter efficiency, long-horizon stability, and cross-resolution generalization, and (iii) factorization reflects a general computational principle for heterogeneous physical systems.

Significance. If the empirical improvements hold, the paper makes a useful architectural contribution: a simple spectral operator that combines a dynamic convolution branch with a static template scaled by global input integrals, while reducing parameter counts and improving stability across several benchmarks. The breadth of the evaluation—spherical, Euclidean, and manifold domains; autoregressive rollout; cross-resolution and cross-regime transfer—is a strength, as is the public code and data availability and the explicit effort to use controlled backbone comparisons. However, the mechanistic claim that the branches 'spontaneously specialize' is not supported by the architecture: the persistent branch is constrained by Eq. (16) to produce a fixed spatial template scaled by a scalar, so the observed specialization is largely constructed rather than discovered. The general 'dynamic–persistent decomposition' claim also outruns the expressive limits of Eq. (16). These issues affect the interpretation, not necessarily the raw accuracy numbers, and can be addressed by reframing the claims and adding targeted ablations.

major comments (4)
  1. [§2.2, Eq. (16)] The 'spontaneous specialization' finding is forced by the persistent-branch definition. For a fixed output channel, S_I = F^{-1}[C_f G_I^θ] is a learned spectral template multiplied by the scalar C_f, the global integral of f. The spatial pattern of S_I therefore cannot vary with the input's spatial distribution; its low across-sample/time variability is a direct consequence of the construction, not a behavior discovered by training. The branch-suppression analysis in Fig. 2 cannot distinguish 'specialization' from architectural constraint. I recommend either removing the word 'spontaneous' and describing the branch as deliberately constrained to a static template, or enriching the persistent branch with input-dependent invariant features and testing whether specialization persists.
  2. [§4.2.2, Eq. (16); §3] The static-template Ansatz limits the claimed general decomposition. A persistent structure whose location, orientation, or support depends on the input—e.g., a downstream wake that follows upstream conditions, or a boundary layer that adapts to local geometry—cannot be represented by C_f G_I^θ, because C_f discards all spatial-distribution information. The full network may compensate through the dynamic branch and channel mixing, but then S_I is effectively a learned bias field, and the mechanistic interpretation in Sec. 2.2 is overstated. The Discussion's caveat about 'richer response decompositions' does not address this, since Eq. (16) already cannot represent input-dependent persistent structure.
  3. [§4.2.1, Eq. (8); SI Eq. (29)] The spherical dynamic-operator derivation assumes a zonal kernel. Eq. (29) integrates over Wigner D-matrices and ends with SHT[G](l,0), which is valid only when the kernel is independent of order m (axisymmetric). The paper does not state this restriction, and Eq. (10) writes G_E^θ as generic learnable spectral weights. Either restrict the theory/notation to zonal kernels on the sphere, or provide the non-zonal convolution derivation and clarify what the learnable weights may depend on.
  4. [§2.1.2, §4.3, Tables 2–3] The controlled architectural comparison is confounded by branch width. FaNO splits the nominal embedding dimension in a 0.5:0.5 channel ratio before concatenation, so its dynamic branch uses roughly half the spectral channels of the matched SFNO/FNO/NORM baseline for the same total embedding dimension, while total parameter counts also differ. The reported gains are therefore not cleanly attributable solely to 'factorization'; reduced dynamic-branch width could itself act as a regularizer. Please add a single-branch ablation at the same total parameter/width budget, or a same-width factorized ablation, to isolate the effect of response factorization.
minor comments (4)
  1. [Eqs. (5), (12)] Eq. (5) integrates over the group G with an invariant measure, but on the sphere v = R u0 is not one-to-one; the stabilizer of u0 should be factored out (G/Stab(u0)) or a normalization convention stated. Eq. (12) appears to mix G and D as integration domains: the notation should be ∫_D G^I(u) f(v) dV(v), not ∫_G.
  2. [Eqs. (13)–(15), Sec. 4.3] The notation for C_f and G_I^θ is inconsistent across domains (k vs. l,m vs. i, and scalar vs. channel-wise). Please specify the dimensions of C_f, the shape of G_I^θ, and whether the integral is computed with quadrature weights or as a plain sum. This matters for interpreting cross-resolution transfer.
  3. [§2.4, Tables 5–6] The text says FaNO achieves the highest ACC at medium- and long-range horizons and remains highly competitive at short range. That is accurate, but at 1-day FaNO's average ACC is below ClimaX/FourCastNet at 32×64 (91.8 vs. 93.0/92.4) and below ClimaX at 64×128 (93.7 vs. 94.2). 'Across all evaluated variables' should be qualified to avoid overstatement.
  4. [Fig. 2 caption, §2.2] The variability statistics for S_I should be interpreted in light of Eq. (16). As written, the text reports low S_I variability as evidence of physical role separation, but that variability is controlled by C_f and the fixed template. Please add a sentence noting this architectural constraint.

Circularity Check

1 steps flagged

Persistent-branch 'spontaneous specialization' is built into Eq. (16): S_I is a fixed spectral template scaled by the global integral C_f, so its low variability and fixed spatial pattern are true by construction.

specific steps
  1. self definitional [Eq. (16), Sec. 4.2.2; interpreted in Sec. 2.2 (Fig. 2) and Sec. 3]
    "g(u) =F −1[Cf ·G I θ],(16) where C f denotes the geometry-aware channel-wise spatial integral of the input function f(u)... Rather than attempting to learn invariance from data, the invariant operators impose it by construction. ... the persistent branch S I produces lower-amplitude but spatially coherent structures whose patterns vary only weakly across samples and time."

    From Eq. (16), for a fixed channel the persistent output is C_f times the fixed learned template F^{-1}[G_I^θ]. Its spatial pattern is therefore identical across all samples up to the global scalar C_f, and it cannot depend on the input's spatial distribution. Consequently, the paper's central mechanistic observation—that S_I has substantially lower sample/time variability and encodes stable, geometry-anchored structures—is a direct mathematical consequence of the definition, not a spontaneous or discovered specialization. The claim that the two branches 'spontaneously specialize into distinct physical roles' (Sec. 3) is imposed by construction; the empirical accuracy gains are independent, but the interpretability/decomposition claim reduces to Eq. (16).

full rationale

The paper's accuracy, parameter-efficiency, cross-resolution, and long-horizon rollout claims are self-contained empirical comparisons against standard baselines (SFNO, FNO, NORM, U-Net, etc.) on public benchmarks, so those results are not circular. However, the central mechanistic claim—that the two branches 'spontaneously specialize' into dynamic vs persistent responses—is forced by Eq. (16). For a fixed channel, S_I = C_f · F^{-1}[G_I^θ]; the spatial pattern is a static template scaled by the input's global integral. Thus Fig. 2's finding that S_I shows low variability and fixed coherent structure is a mathematical consequence of the architecture, not evidence of learned emergence. The Discussion's caveat about 'richer response decompositions' does not resolve this: the static-template restriction already limits the targeted dynamic–persistent decomposition. The mechanistic specialization is therefore partially circular (score 6), while the empirical improvements remain valid and would stand even if the specialization claim were removed.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The central claim rests on two learned spectral kernels (dynamic and persistent), a channel-mixing layer, and the persistent-branch Ansatz that persistent structure equals C_f times a static template. The Ansatz is an ad-hoc modeling assumption, not derived from physics. The standard spectral convolution theorems are external background.

free parameters (4)
  • G_E_theta (dynamic spectral kernel) = learned during training
    Fitted to data via training; appears in Eqs (10), (17)-(20) as the equivariant filter.
  • G_I_theta (persistent spectral template) = learned during training
    Fitted to data; Eq (16). It defines a static spatial template whose only input dependence is the scalar C_f.
  • w1 (channel-mixing linear layer) = learned during training
    Fitted; Eq (17) after concatenation of the two branches.
  • branch channel ratio = 0.5 : 0.5
    Hand-chosen default in Sec. 4.3; affects expressivity allocation and parameter counts.
axioms (6)
  • standard math Green's function solution formula for linear PDEs (Eqs 1-4)
    Assumes the target system is linear and the Green's function integral gives the solution.
  • standard math Spectral convolution theorem on Euclidean domains (Eq 7)
    Translational equivariance implies Fourier diagonalization.
  • domain assumption Spherical convolution theorem restricted to zonal filters (Eq 8: G_E(l,0))
    The derivation only yields a clean rotation-equivariant product for zonal kernels; non-zonal dynamics are excluded without discussion.
  • domain assumption Manifold kernels are diagonal in the Laplace–Beltrami eigenbasis (Eqs 9, 36)
    This holds for operators that commute with the Laplacian, not for general Green's functions; the paper treats it as a design choice.
  • ad hoc to paper Persistent physical structure can be represented by an absolute-position kernel G_I(u)
    The central persistent-branch Ansatz (Eqs 11,16) is a modeling postulate, not derived from the PDE or data.
  • ad hoc to paper Branches spontaneously specialize without explicit supervision
    The claimed mechanism in Sec. 2.2 is an interpretation; the architecture itself forces the persistent branch to be low-variability.
invented entities (1)
  • Persistent response component S_I no independent evidence
    purpose: Encodes stable geometry-induced spatial templates that are intended to carry invariant, slowly varying structure.
    Introduced by the architecture (Eq. 16); no external falsifiable prediction outside the paper. Its low variability is a consequence of the C_f-scaled static template, so it is an invented modeling construct rather than a measured physical quantity.

pith-pipeline@v1.3.0-alltime-deepseek · 26429 in / 15364 out tokens · 158411 ms · 2026-08-02T11:06:46.545800+00:00 · methodology

0 comments
read the original abstract

Physical systems often exhibit heterogeneous mechanisms, where rapidly evolving dynamics coexist with persistent structures. Capturing such multiscale physical behavior remains challenging for existing neural operators, which typically rely on single dominant inductive bias and therefore couple distinct physical responses into a shared representation. We introduce the Unified Green's Function Framework across domains and propose the Factorized Neural Operators (FaNO), which decompose spectral representations into equivariant-inspired dynamic responses and invariant-inspired persistent responses, leading to better interpretability and generalization. Mechanistically, we show that the two operator branches spontaneously specialize into distinct physical roles that remain consistent across scales and domains: the equivariant-inspired branch captures rapidly varying transient dynamics, whereas the invariant-inspired branch extracts coherent persistent structures. This factorized mechanism of FaNO consistently improves prediction accuracy, parameter efficiency and cross-scale generalization across physical systems and domains. In particular, it maintains consistent predictions under long-horizon autoregressive rollout, cross-resolution extrapolation and physical-regime shifts. These findings suggest that scalable physical modeling may benefit from moving beyond single-inductive-bias formulations toward factorized operator representations that better reflect the heterogeneous organization of physical systems, accelerating the reliable deployment of machine learning for scientific computing and discovery.

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 8 linked inside Pith

  1. [1]

    In: International Conference on Machine Learning, pp

    Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., Anand- kumar, A.: Spherical fourier neural operators: Learning stable dynamics on the sphere. In: International Conference on Machine Learning, pp. 2806–2823 (2023). PMLR

    2023

  2. [2]

    npj Climate and Atmospheric Science, 7 (1), 221 (2024)

    Liu, C., Hsu, K., Peng, M., Chen, D., Chang, P., Hsiao, L., et al.: Evaluation of five global AI models for predicting weather in Eastern Asia and Western Pacific. npj Climate and Atmospheric Science, 7 (1), 221 (2024)

    2024

  3. [3]

    Scientific Reports15(1), 11522 (2025)

    Hu, Y., Yin, F., Zhang, W., Ren, K., Song, J., Deng, K., Zhang, D.: Spherical multigrid neural operator for improving autoregressive global weather forecasting. Scientific Reports15(1), 11522 (2025)

    2025

  4. [4]

    In: The Fourteenth International Conference on Learning Representations (2026).https://openreview.net/forum?id=XkGjzSDTnm

    Tang, H., Chen, H., Li, C.: Generalized spherical neural operators: Green’s function formulation. In: The Fourteenth International Conference on Learning Representations (2026).https://openreview.net/forum?id=XkGjzSDTnm

    2026

  5. [5]

    IEEE transactions on medical imaging40(4), 1217–1228 (2021)

    Zhao, F., Wu, Z., Wang, L., Lin, W., Gilmore, J.H., Xia, S., Shen, D., Li, G.: Spherical deformable u-net: Application to cortical surface parcellation and development prediction. IEEE transactions on medical imaging40(4), 1217–1228 (2021)

    2021

  6. [6]

    Communications in computational physics2(3), 367–450 (2007)

    Weinan, E., Engquist, B., Li, X., Ren, W., Vanden-Eijnden, E.: Heterogeneous multiscale methods: a review. Communications in computational physics2(3), 367–450 (2007)

    2007

  7. [7]

    Archives of Computational Methods in Engineering28(3), 1017–1037 (2021)

    Peng, G.C., Alber, M., Buganza Tepole, A., Cannon, W.R., De, S., Dura-Bernal, S., Garikipati, K., Karniadakis, G., Lytton, W.W., Perdikaris, P.,et al.: Multiscale modeling meets machine learning: What can we learn? gcy peng et al. Archives of Computational Methods in Engineering28(3), 1017–1037 (2021)

    2021

  8. [8]

    Physical Review Letters133(24), 244201 (2024)

    Lucarini, V., Chekroun, M.D.: Detecting and attributing change in climate and complex systems: Foundations, green’s functions, and nonlinear fingerprints. Physical Review Letters133(24), 244201 (2024)

    2024

  9. [9]

    In: The Thirteenth International Conference on Learning Representations (2025)

    Gao, W., Xu, R., Deng, Y., Liu, Y.: Discretization-invariance? on the dis- cretization mismatch errors in neural operators. In: The Thirteenth International Conference on Learning Representations (2025)

    2025

  10. [10]

    arXiv preprint arXiv:2003.03485 (2020)

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

    Pith/arXiv arXiv 2003

  11. [11]

    Journal of Machine Learning Research24(89), 1–97 (2023)

    Kovachki, N., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A., Anandkumar, A.: Neural operator: Learning maps between function spaces with 22 applications to pdes. Journal of Machine Learning Research24(89), 1–97 (2023)

    2023

  12. [12]

    Proceedings of the IEEE86(11), 2278–2324 (1998)

    LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. Proceedings of the IEEE86(11), 2278–2324 (1998)

    1998

  13. [13]

    In: International Conference on Medical Image Computing and Computer-assisted Intervention, pp

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedical image segmentation. In: International Conference on Medical Image Computing and Computer-assisted Intervention, pp. 234–241 (2015). Springer

    2015

  14. [14]

    Neural networks2(5), 359–366 (1989)

    Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal approximators. Neural networks2(5), 359–366 (1989)

    1989

  15. [15]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recogni- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016)

    2016

  16. [16]

    arXiv preprint arXiv:2207.11417 (2022)

    L¨ utjens, B., Crawford, C.H., Watson, C.D., Hill, C., Newman, D.: Multiscale neural operator: Learning fast and grid-independent pde solvers. arXiv preprint arXiv:2207.11417 (2022)

    Pith/arXiv arXiv 2022

  17. [17]

    arXiv preprint arXiv:2010.08895 (2020)

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 (2020)

    Pith/arXiv arXiv 2010

  18. [18]

    Nature communications12(1), 1330 (2021)

    Giambagli, L., Buffoni, L., Carletti, T., Nocentini, W., Fanelli, D.: Machine learning in spectral domain. Nature communications12(1), 1330 (2021)

    2021

  19. [19]

    National Science Open3(6), 20240001 (2024)

    Chen, G., Liu, X., Meng, Q., Chen, L., Liu, C., Li, Y.: Learning neural operators on riemannian manifolds. National Science Open3(6), 20240001 (2024)

    2024

  20. [20]

    arXiv preprint arXiv:2202.07230 (2022)

    Han, J., Rong, Y., Xu, T., Huang, W.: Geometrically equivariant graph neural networks: A survey. arXiv preprint arXiv:2202.07230 (2022)

    Pith/arXiv arXiv 2022

  21. [21]

    Advances in Neural Information Processing Systems34, 30037–30049 (2021)

    Finzi, M., Benton, G., Wilson, A.G.: Residual pathway priors for soft equivariance constraints. Advances in Neural Information Processing Systems34, 30037–30049 (2021)

    2021

  22. [22]

    Advances in Neural Information Processing Systems32(2019)

    Worrall, D., Welling, M.: Deep scale-spaces: Equivariance over scale. Advances in Neural Information Processing Systems32(2019)

    2019

  23. [23]

    In: Interna- tional Conference on Machine Learning, pp

    Cohen, T., Welling, M.: Group equivariant convolutional networks. In: Interna- tional Conference on Machine Learning, pp. 2990–2999 (2016). PMLR

    2016

  24. [24]

    In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp

    Zheng, Z., Liu, Y., Li, J., Yao, J., Rong, Y.: Relaxing continuous constraints of equivariant graph neural networks for broad physical dynamics learning. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 4548–4558 (2024) 23

    2024

  25. [25]

    In: International Conference on Machine Learning, pp

    Wang, R., Walters, R., Yu, R.: Approximately equivariant networks for imper- fectly symmetric dynamics. In: International Conference on Machine Learning, pp. 23078–23091 (2022). PMLR

    2022

  26. [26]

    In: European Conference on Computer Vision, pp

    Wad, T., Sun, Q., Pranata, S., Jayashree, K., Zhang, H.: Equivariance and invari- ance inductive bias for learning from insufficient data. In: European Conference on Computer Vision, pp. 241–258 (2022). Springer

    2022

  27. [27]

    Computer Physics Reports1(2), 57–125 (1984)

    Niessen, W., Schirmer, J., Cederbaum, L.S.: Computational methods for the one- particle green’s function. Computer Physics Reports1(2), 57–125 (1984)

    1984

  28. [28]

    Reports on Progress in Physics 82(10), 106801 (2019)

    Pan, E.: Green’s functions for geophysics: a review. Reports on Progress in Physics 82(10), 106801 (2019)

    2019

  29. [29]

    Philosophy of science1(2), 163–169 (1934)

    Einstein, A.: On the method of theoretical physics. Philosophy of science1(2), 163–169 (1934)

    1934

  30. [30]

    Journal of Computational Physics362, 425–448 (2018)

    Bonev, B., Hesthaven, J.S., Giraldo, F.X., Kopera, M.A.: Discontinuous galerkin scheme for the spherical shallow water equations with applications to tsunami modeling and prediction. Journal of Computational Physics362, 425–448 (2018)

    2018

  31. [31]

    International Journal for Numerical Methods in Fluids35(8), 869–901 (2001)

    Giraldo, F.X.: A spectral element shallow water model on spherical geodesic grids. International Journal for Numerical Methods in Fluids35(8), 869–901 (2001)

    2001

  32. [32]

    arXiv preprint arXiv:2301.10343 (2023)

    Nguyen, T., Brandstetter, J., Kapoor, A., Gupta, J.K., Grover, A.: Climax: A foundation model for weather and climate. arXiv preprint arXiv:2301.10343 (2023)

    Pith/arXiv arXiv 2023

  33. [33]

    arXiv preprint arXiv:2202.11214 (2022)

    Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mar- dani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., et al.: Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators. arXiv preprint arXiv:2202.11214 (2022)

    Pith/arXiv arXiv 2022

  34. [34]

    Journal of Advances in Modeling Earth Systems12(11), 2020–002203 (2020)

    Rasp, S., Dueben, P.D., Scher, S., Weyn, J.A., Mouatadid, S., Thuerey, N.: Weath- erbench: a benchmark data set for data-driven weather forecasting. Journal of Advances in Modeling Earth Systems12(11), 2020–002203 (2020)

    2020

  35. [35]

    Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020)

    Hersbach, H., Bell, B., Berrisford, P., Hirahara, S., Hor´ anyi, A., Mu˜ noz-Sabater, J., Nicolas, J., Peubey, C., Radu, R., Schepers, D.,et al.: The era5 global reanal- ysis. Quarterly journal of the royal meteorological society146(730), 1999–2049 (2020)

    1999

  36. [36]

    In: International Conference on Learning Representations (2021)

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh- based simulation with graph networks. In: International Conference on Learning Representations (2021)

    2021

  37. [37]

    ACM Transactions on Graphics (TOG)41(3), 1–16 (2022)

    Sharp, N., Attaiki, S., Crane, K., Ovsjanikov, M.: Diffusionnet: Discretization 24 agnostic learning on surfaces. ACM Transactions on Graphics (TOG)41(3), 1–16 (2022)

    2022

  38. [38]

    Advances in Neural Information Processing Systems36, 71242–71262 (2023)

    Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M.W., Gholami, A.: Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. Advances in Neural Information Processing Systems36, 71242–71262 (2023)

    2023

  39. [39]

    arXiv preprint arXiv:2301.12664 (2023)

    Wu, H., Hu, T., Luo, H., Wang, J., Long, M.: Solving high-dimensional pdes with latent spectral models. arXiv preprint arXiv:2301.12664 (2023)

    Pith/arXiv arXiv 2023

  40. [40]

    Journal of Computational Physics, 114530 (2025)

    You, Z., Xu, Z., Cai, W.: Mscalefno: Multi-scale fourier neural operator learning for oscillatory functions and wave scattering problems. Journal of Computational Physics, 114530 (2025)

    2025

  41. [41]

    arXiv preprint arXiv:1801.10130 (2018)

    Cohen, T.S., Geiger, M., K¨ ohler, J., Welling, M.: Spherical cnns. arXiv preprint arXiv:1801.10130 (2018)

    Pith/arXiv arXiv 2018

  42. [42]

    In: International Conference on Machine Learning, pp

    Hao, Z., Wang, Z., Su, H., Ying, C., Dong, Y., Liu, S., Cheng, Z., Song, J., Zhu, J.: Gnot: A general neural operator transformer for operator learning. In: International Conference on Machine Learning, pp. 12556–12569 (2023). PMLR

    2023

  43. [43]

    Advances in applied mathematics15(2), 202–250 (1994)

    Driscoll, J.R., Healy, D.M.: Computing fourier transforms and convolutions on the 2-sphere. Advances in applied mathematics15(2), 202–250 (1994)

    1994

  44. [44]

    M¨ uller, C.: Spherical Harmonics vol. 17. Springer, Berlin (2006)

    2006

  45. [45]

    arXiv preprint arXiv:2512.16409 (2025)

    Tang, H., Zhu, J., Feng, Z., Li, H., Li, C.: Geometric laplace neural operator. arXiv preprint arXiv:2512.16409 (2025)

    arXiv 2025

  46. [46]

    IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    Bronstein, M.M., Bruna, J., LeCun, Y., Szlam, A., Vandergheynst, P.: Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    2017

  47. [47]

    1 CV CX c=1 VX v=1 σ(t) c,v 2 #1/2 , t= 1, . . . , T.(54) Equivalently, V (t) sample =

    Poulenard, A., Rakotosaona, M.-J., Ponty, Y., Ovsjanikov, M.: Effective rotation- invariant point cnn with spherical harmonics kernels. In: 2019 International Conference on 3D Vision (3DV), pp. 47–56 (2019). IEEE 25 Supplementary Information A Detailed derivation of factorized neural operators A.1 Dynamic operator derivation Detailed derivations of dynami...

    arXiv 2019