arxiv: 2511.22112 · v1 · submitted 2025-11-27 · 💻 cs.LG

Recognition: no theorem link

Toward Data-Driven Surrogates of the Solar Wind with Spherical Fourier Neural Operator

Reza Mansouri , Dustin Kempton , Pete Riley , Rafal Angryk

Authors on Pith no claims yet

Pith reviewed 2026-05-17 04:24 UTC · model grok-4.3

classification 💻 cs.LG

keywords solar windspherical fourier neural operatorsurrogate modelingspace weatherneural operatorsdata-driven forecastingheliophysics

0 comments

The pith

A Spherical Fourier Neural Operator delivers a data-driven surrogate for steady-state solar wind modeling that matches or exceeds HUX performance on standard metrics.

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

The paper introduces a Spherical Fourier Neural Operator as a trainable surrogate to simulate the steady-state solar wind, which carries charged particles from the Sun and drives space weather effects on Earth. Current three-dimensional MHD simulations are accurate but too slow for rapid exploration of boundary uncertainties or real-time use. The authors train and test the SFNO on solar wind velocity data and compare it directly to the established HUX numerical surrogate. Across several scalar error and correlation metrics the SFNO performs at least as well, and sometimes better, than HUX. The work notes that HUX still produces smoother, more physically consistent fields and therefore calls for new evaluation standards that better capture physical fidelity.

Core claim

The central claim is that a Spherical Fourier Neural Operator, trained on solar wind simulation data, produces steady-state velocity fields whose accuracy on standard metrics is comparable to or better than the HUX numerical surrogate while remaining computationally efficient and improvable with additional training data.

What carries the argument

The Spherical Fourier Neural Operator (SFNO), a neural-network architecture that learns direct mappings from boundary conditions to output velocity fields on the sphere.

If this is right

SFNO enables rapid evaluation of many boundary-condition scenarios that would be too expensive to run with full MHD codes.
Trained models support real-time or near-real-time solar wind forecasting once inference speed is achieved.
Accuracy can continue to rise as larger archives of solar wind observations or simulations become available for training.
The same architecture can be retrained or fine-tuned for related heliospheric quantities such as magnetic field or density.

Where Pith is reading between the lines

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

SFNO-style operators could be extended to transient events such as coronal mass ejections by adding time as an input dimension.
Ensemble forecasts that combine SFNO speed with occasional full MHD runs become practical for operational space-weather centers.
Similar spherical neural operators may transfer to other global spherical problems such as atmospheric circulation or ocean modeling.

Load-bearing premise

That the chosen scalar error and correlation metrics are sufficient to judge whether the predicted solar wind fields are physically realistic.

What would settle it

A side-by-side check of velocity-field smoothness, divergence, or conservation properties in SFNO outputs versus HUX and full MHD runs, where large unphysical artifacts appear only in the SFNO fields.

Figures

Figures reproduced from arXiv: 2511.22112 by Dustin Kempton, Pete Riley, Rafal Angryk, Reza Mansouri.

**Figure 2.** Figure 2: Equiangular projections of solar wind radial velocity in one instance [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Figure 3: Solar wind radial velocity at R ≈ 49 R⊙ for Carrington Rotation 2293 from the MHD solution. The edge regions, detected using a Sobel filter, are outlined with dashed lines. D. Training Strategy 1) Loss Function and Optimization: The model is trained using a layer-wise two-dimensional L2 loss, defined as L (2D) 2 = 1 BC B X×C b,c   H X×W i,j |ybcij − yˆbcij | 2   1/2 (1) where B is the batch size, C the… view at source ↗

**Figure 4.** Figure 4: Cross-validation results showing the MSE across five folds for SFNO [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

**Figure 5.** Figure 5: Training and test loss curves for the optimal SFNO configuration [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 8.** Figure 8: MSE across radius index (1–139, up to 1 AU) for the SFNO and [PITH_FULL_IMAGE:figures/full_fig_p004_8.png] view at source ↗

**Figure 7.** Figure 7: Distribution of solar wind speed values for Carrington Rotation 2293, [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗

**Figure 9.** Figure 9: Resolution invariance of the optimal 8-layer, 256-channel SFNO model, which operates in the spectral domain using spherical harmonics. Spherical [PITH_FULL_IMAGE:figures/full_fig_p005_9.png] view at source ↗

**Figure 10.** Figure 10: Solar wind speed at 1 AU for Carrington Rotation 2228. (A) Ground [PITH_FULL_IMAGE:figures/full_fig_p005_10.png] view at source ↗

read the original abstract

The solar wind, a continuous stream of charged particles from the Sun's corona, shapes the heliosphere and impacts space systems near Earth. Variations such as high-speed streams and coronal mass ejections can disrupt satellites, power grids, and communications, making accurate modeling essential for space weather forecasting. While 3D magnetohydrodynamic (MHD) models are used to simulate and investigate these variations in the solar wind, they tend to be computationally expensive, limiting their usefulness in investigating the impacts of boundary condition uncertainty. In this work, we develop a surrogate for steady state solar wind modeling, using a Spherical Fourier Neural Operator (SFNO). We compare our model to a previously developed numerical surrogate for this task called HUX, and we show that the SFNO achieves comparable or better performance across several metrics. Though HUX retains advantages in physical smoothness, this underscores the need for improved evaluation criteria rather than a flaw in SFNO. As a flexible and trainable approach, SFNO enables efficient real-time forecasting and can improve with more data. The source code and more visual results are available at https://github.com/rezmansouri/solarwind-sfno-velocity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SFNO gives a workable velocity surrogate for solar wind but the scalar metrics leave physical fidelity unproven.

read the letter

The main point is that this paper trains a Spherical Fourier Neural Operator on simulation data to act as a fast surrogate for steady-state solar wind velocity and then compares it directly to the HUX baseline. They report that SFNO matches or beats HUX on the metrics they chose while being more flexible for future data additions. That comparison and the released code are the concrete new pieces here. Prior work had not shown this head-to-head on the solar-wind task with this architecture. The authors also flag that HUX still looks smoother in a physical sense, which keeps the claim grounded rather than overstated. The evaluation is the clear weak spot. The abstract supplies pointwise numbers without error bars or any description of training and validation splits. The paper itself notes that the chosen metrics may miss smoothness and conservation properties that matter for downstream forecasting or uncertainty work. If the model produces fields that score well on magnitude error but violate approximate divergence-free behavior or create unphysical jumps, the performance numbers do not yet establish it as a reliable replacement for MHD ensembles. This is aimed at heliophysics researchers who already use or build surrogates for space-weather boundary conditions. A reader who wants to see how neural operators handle spherical geometry in this domain will get a usable example and a starting code base. It is not a finished operational tool, but the direct baseline comparison and the authors' own caveat about evaluation criteria make it worth a full referee process rather than a desk reject. I would send it to review and ask specifically for tests that check physical consistency beyond scalar error.

Referee Report

1 major / 0 minor

Summary. The manuscript develops a Spherical Fourier Neural Operator (SFNO) as a data-driven surrogate for steady-state solar wind velocity, trained on external simulation data. It compares SFNO to the HUX numerical baseline and claims that SFNO achieves comparable or better performance across several (scalar) metrics, while noting that HUX retains advantages in physical smoothness and calling for improved evaluation criteria.

Significance. If the evaluation is strengthened with statistical rigor and physical-consistency checks, the work could supply a flexible, trainable surrogate that improves computational efficiency over full MHD models for heliospheric forecasting and uncertainty quantification. The open-source code and emphasis on data-driven improvement are positive features.

major comments (1)

[Abstract] Abstract: performance numbers are reported without error bars, without any description of training/validation splits or cross-validation procedure, and without held-out tests for physical properties such as smoothness or approximate divergence-free behavior. The abstract itself states that the chosen metrics may not capture physical smoothness; this directly undermines the central claim that SFNO constitutes a valid surrogate for solar-wind modeling.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed review and valuable suggestions. We have addressed the major comment on the abstract and evaluation by revising the manuscript to include more rigorous reporting of results and additional discussion on physical consistency.

read point-by-point responses

Referee: [Abstract] Abstract: performance numbers are reported without error bars, without any description of training/validation splits or cross-validation procedure, and without held-out tests for physical properties such as smoothness or approximate divergence-free behavior. The abstract itself states that the chosen metrics may not capture physical smoothness; this directly undermines the central claim that SFNO constitutes a valid surrogate for solar-wind modeling.

Authors: We appreciate the referee pointing out the need for greater statistical rigor in the abstract. The detailed description of the data splits, training procedure, and cross-validation is provided in the Methods section of the manuscript. To address the concern, we have updated the abstract to include error bars on the reported performance metrics and a brief mention of the validation approach. For held-out tests on physical properties: our current evaluation uses standard metrics from the solar wind literature for direct comparison with HUX. We have not conducted explicit tests for approximate divergence-free behavior in this study, as the focus is on velocity field prediction from simulation data. We acknowledge this limitation and have added text in the revised manuscript discussing potential physical consistency checks as future work. The note in the abstract about the metrics not fully capturing physical smoothness is intended to provide balance and does not undermine our claim. The central claim is that the SFNO provides a flexible, data-driven alternative that performs comparably or better on the chosen metrics and can be improved with additional training data. We explicitly recognize HUX's strengths in smoothness to highlight opportunities for better surrogate evaluation criteria. This is consistent with the significance of the work as a trainable surrogate for efficiency in forecasting and uncertainty quantification. revision: yes

Circularity Check

0 steps flagged

No significant circularity in SFNO solar wind surrogate claims

full rationale

The paper trains a Spherical Fourier Neural Operator on external MHD simulation data to produce steady-state solar wind velocity fields and evaluates it against the independent HUX numerical surrogate using standard scalar metrics. No load-bearing step in the described workflow reduces by construction to a fitted parameter, self-definition, or self-citation chain; the central performance comparison rests on held-out simulation data and an externally developed baseline. The manuscript itself flags HUX's retained advantages in physical smoothness, confirming that evaluation choices are presented as provisional rather than internally forced.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the neural operator can be trained to reproduce steady-state MHD outputs and that the reported scalar metrics are adequate proxies for physical fidelity. No new physical entities are postulated.

free parameters (1)

SFNO network weights
Learned during training on simulation data; exact count and regularization not stated in abstract.

axioms (1)

domain assumption Steady-state solar wind can be adequately represented by velocity fields on a spherical grid
Implicit in the choice of SFNO architecture and comparison to HUX.

pith-pipeline@v0.9.0 · 5510 in / 1104 out tokens · 21437 ms · 2026-05-17T04:24:19.648818+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

[1]

Space technology 5,

NASA, “Space technology 5,” Jan 2025. [Online]. Available: https://www.jpl.nasa.gov/nmp/st5/SCIENCE/solarwind.html

work page 2025
[2]

National Academies Press, dec 2008

National Research Council,Severe Space Weather Events– Understanding Societal and Economic Impacts. National Academies Press, dec 2008. [Online]. Available: https://doi.org/10.17226/12507

work page doi:10.17226/12507 2008
[3]

Description of a singular appearance seen in the sun on September 1, 1859,

R. C. Carrington, “Description of a singular appearance seen in the sun on September 1, 1859,”Mon. Not. R. Astron. Soc., vol. 20, no. 1, pp. 13– 15, nov 1859. [Online]. Available: https://doi.org/10.1093/mnras/20.1.13

work page doi:10.1093/mnras/20.1.13
[4]

Large-scale structure of the interplanetary medium,

J. T. Nolte and E. C. Roelof, “Large-scale structure of the interplanetary medium,”Solar Physics, vol. 33, no. 1, pp. 241–257, Nov 1973. [Online]. Available: https://doi.org/10.1007/BF00152395

work page doi:10.1007/bf00152395 1973
[5]

Spatial structure of the solar wind and comparisons with solar data and models,

M. Neugebauer, R. J. Forsyth, A. B. Galvin, K. L. Harvey, J. T. Hoeksema, A. J. Lazarus, R. P. Lepping, J. A. Linker, Z. Mikic, J. T. Steinberg, R. von Steiger, Y .-M. Wang, and R. F. Wimmer- Schweingruber, “Spatial structure of the solar wind and comparisons with solar data and models,”Journal of Geophysical Research: Space Physics, vol. 103, no. A7, pp....

work page doi:10.1029/98ja00798 1998
[6]

Improvement in the prediction of solar wind conditions using real-time solar magnetic field updates,

C. Arge and V . Pizzo, “Improvement in the prediction of solar wind conditions using real-time solar magnetic field updates,”Journal of Geophysical Research, vol. 105, pp. 10 465–10 480, 05 2000

work page 2000
[7]

Mapping solar wind streams from the sun to 1 au: A comparison of techniques,

P. Riley and R. Lionello, “Mapping solar wind streams from the sun to 1 au: A comparison of techniques,”Solar Physics, vol. 270, no. 2, pp. 575–592, Jun 2011. [Online]. Available: https://doi.org/10.1007/s11207-011-9766-x

work page doi:10.1007/s11207-011-9766-x 2011
[8]

Using a heliospheric upwinding extrapolation technique to magnetically connect different regions of the heliosphere,

P. Riley and O. Issan, “Using a heliospheric upwinding extrapolation technique to magnetically connect different regions of the heliosphere,” Frontiers in Physics, vol. V olume 9 - 2021, 2021. [Online]. Available: https://www.frontiersin.org/journals/physics/articles/10.3389/ fphy.2021.679497

work page arXiv 2021
[9]

An empirically-driven global mhd model of the solar corona and inner heliosphere,

P. Riley, J. A. Linker, and Z. Miki ´c, “An empirically-driven global mhd model of the solar corona and inner heliosphere,”Journal of Geophysical Research: Space Physics, vol. 106, no. A8, pp. 15 889–15 901, 2001. [Online]. Available: https://agupubs.onlinelibrary. wiley.com/doi/abs/10.1029/2000JA000121

work page doi:10.1029/2000ja000121 2001
[10]

Gramacy,Surrogates: Gaussian Process Modeling, Design, and Optimization for the Applied Sciences, ser

R. Gramacy,Surrogates: Gaussian Process Modeling, Design, and Optimization for the Applied Sciences, ser. Chapman & Hall/CRC Texts in Statistical Science. CRC Press, 2020. [Online]. Available: https://books.google.com/books?id=1w WDwAAQBAJ

work page 2020
[11]

Ensemble Modeling of CMEs Using the WSA–ENLIL+Cone Model,

M. L. Mays, A. Taktakishvili, A. Pulkkinen, P. J. MacNeice, L. Rast ¨atter, D. Odstrcil, L. K. Jian, I. G. Richardson, J. A. LaSota, Y . Zheng, and M. M. Kuznetsova, “Ensemble Modeling of CMEs Using the WSA–ENLIL+Cone Model,”Space Physics, vol. 290, no. 6, pp. 1775–1814, Jun. 2015. [Online]. Available: https://doi.org/10.1007/s11207-015-0692-1

work page doi:10.1007/s11207-015-0692-1 2015
[12]

Corotating interaction regions during the recent solar minimum: The power and limitations of global mhd modeling,

P. Riley, J. A. Linker, R. Lionello, and Z. Mikic, “Corotating interaction regions during the recent solar minimum: The power and limitations of global mhd modeling,”Journal of Atmospheric and Solar-Terrestrial Physics, vol. 83, pp. 1–10, 2012, corotating Interaction Regions from Sun to Earth: Modeling their formation, evolution and geoeffectiveness. [Onl...

work page doi:10.1016/j.jastp.2011.12.013 2012
[13]

Spherical fourier neural operators: Learning stable dynamics on the sphere,

B. Bonev, T. Kurth, C. Hundt, J. Pathak, M. Baust, K. Kashinath, and A. Anandkumar, “Spherical fourier neural operators: Learning stable dynamics on the sphere,” 2023. [Online]. Available: https: //arxiv.org/abs/2306.03838

work page arXiv 2023
[14]

Solar wind prediction using deep learning,

V . Upendran, M. C. M. Cheung, S. Hanasoge, and G. Krishnamurthi, “Solar wind prediction using deep learning,” Space Weather, vol. 18, no. 9, p. e2020SW002478, 2020, e2020SW002478 10.1029/2020SW002478. [Online]. Available: https: //agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2020SW002478

work page doi:10.1029/2020sw002478 2020
[15]

Swan: A multihead autoregressive attention model for solar wind speed forecasting,

M. Cobos-Maestre, M. Flores-Soriano, and D. F. Barrero, “Swan: A multihead autoregressive attention model for solar wind speed forecasting,”Expert Systems with Applications, vol. 252, p. 124128, 2024. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0957417424009941

work page 2024
[16]

Operator learning: Algorithms and analysis,

N. B. Kovachki, S. Lanthaler, and A. M. Stuart, “Operator learning: Algorithms and analysis,” 2024. [Online]. Available: https: //arxiv.org/abs/2402.15715

work page arXiv 2024
[17]

Fourier Neural Operator for Parametric Partial Differential Equations

Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, “Fourier neural operator for parametric partial differential equations,” 2021. [Online]. Available: https://arxiv.org/abs/2010.08895

work page internal anchor Pith review Pith/arXiv arXiv 2021
[18]

Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators,

J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopadhyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, P. Hassanzadeh, K. Kashinath, and A. Anandkumar, “Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators,” 2022. [Online]. Available: https://arxiv.org/abs/2202. 11214

work page 2022
[19]

The tilts of corotating interaction regions at midheliographic latitudes,

P. Riley, J. T. Gosling, L. A. Weiss, and V . J. Pizzo, “The tilts of corotating interaction regions at midheliographic latitudes,”Journal of Geophysical Research: Space Physics, vol. 101, no. A11, pp. 24 349–24 357, 1996. [Online]. Available: https://agupubs.onlinelibrary. wiley.com/doi/abs/10.1029/96JA02447

work page doi:10.1029/96ja02447 1996
[20]

Multiscale structural similarity for image quality assessment,

Z. Wang, E. Simoncelli, and A. Bovik, “Multiscale structural similarity for image quality assessment,” inThe Thrity-Seventh Asilomar Confer- ence on Signals, Systems & Computers, 2003, vol. 2, 2003, pp. 1398– 1402 V ol.2

work page 2003
[21]

A library for learning neural operators,

J. Kossaifi, N. Kovachki, Z. Li, D. Pitt, M. Liu-Schiaffini, R. J. George, B. Bonev, K. Azizzadenesheli, J. Berner, and A. Anandkumar, “A library for learning neural operators,” 2024

work page 2024
[22]

Neural operator: Learning maps between function spaces,

N. B. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. M. Stuart, and A. Anandkumar, “Neural operator: Learning maps between function spaces,”CoRR, vol. abs/2108.08481, 2021

work page arXiv 2021