Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Modeling turbulent and self-gravitating fluids with Fourier neural operators

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fourier neural operators trained only on projected 2D snapshots can forecast the future of collapsing and turbulent gas, even when a magnetic field is hidden from the inputs.

desk verdict Solid proof-of-concept for learned surrogates of projected astrophysical flows, but the central claim about predicting unobserved variables outruns the experiments. read the letter →

arxiv 2507.23662 v1 pith:TKYVDUGP submitted 2025-07-31 astro-ph.GA astro-ph.IM

Keith Poletti , Stella S. R. Offner , Rachel A. Ward This is my paper
classification astro-ph.GAastro-ph.IM
keywords Fourierneuraloperatorssurrogatesmagnetohydrodynamicturbulencesupersonicgravitationalcollapseobservationalprojectioncolumndensitysurrogatemodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether Fourier neural operators—networks that learn a map between function spaces rather than a single function—can forecast the evolution of astrophysical gas from the same kind of incomplete data telescopes actually provide: a two-dimensional column density image and a plane-of-sky velocity field. The authors train these networks on numerical simulations of a gravitationally collapsing sphere, supersonic hydrodynamic turbulence, and supersonic magnetohydrodynamic turbulence, always projecting the 3D state into 2D observational proxies. They report that the trained operators predict future projected snapshots with normalized root-mean-square errors of roughly 0.05–0.25, covering a density dynamic range of four orders of magnitude. The most striking claim is that the networks can predict the effects of a dynamical variable—the magnetic field—that is never included in the training inputs. If true, neural surrogates could someday forecast direct astronomical observables without needing a complete 3D physical state.

What carries the argument

The load-bearing object is the projection φ that sends the 3D simulation state to a 2D observational proxy—column density and plane-of-sky velocity—together with the assumed operator G on the observation space that forward-propagates these proxies in time. The neural networks (FNO-3D, autoregressive FNO-3D, and a U-shaped neural operator) approximate G by stacking Fourier layers that convolve the input in the spectral domain and truncate high-frequency modes. The projection is non-invertible, so the entire approach depends on G being well-defined on the low-dimensional space; the paper states this assumption explicitly and approximates G by minimizing a relative L2 error between predicted and true projected outputs.

What would settle it

Run the MHD experiment with varied initial magnetic field orientations while keeping the projected density and velocity inputs fixed; if two runs with nearly identical projections evolve to visibly different projected futures, the fixed operator G on observation space does not exist and the surrogate's errors should grow correspondingly.

Watch

Extended reading notes

Core claim

The central claim is that a nonlinear operator acting directly on the projected observation space exists and can be learned: given the sequence (σ, ū, v̄) of column density and plane-of-sky velocity over four timesteps, the network maps to the same observables roughly fifteen timesteps later, matching the simulated truth better than the identity baseline in every experiment. For spherical collapse the method tracks the growth of the central core as the projected density spans four orders of magnitude. For supersonic turbulence (Mach 12), the models reproduce large-scale structure while smearing small filaments, and the UNO variant preserves the most small-scale power. For magnetized turbulence, omitting the magnetic field from the inputs degrades accuracy relative to full-information settings, but the networks still outperform the identity baseline, which the authors take as evidence that FNOs capture the effect of unobserved dynamical variables.

Load-bearing premise

The approach assumes that a single fixed map can advance the 2D projected images in time, even though many different 3D states can project to the same image; if hidden variables like the magnetic field steer the future in ways the projection does not record, that map may not exist.

Editorial extensions

If this is right

  • If the learned operators generalize to unseen initial states, neural surrogates could reduce the search space for inverse problems in astrophysics, with high-fidelity PDE solvers used only to refine the final answer.
  • Forecasting in observational space sidesteps the need to reconstruct the full 3D state, which telescope data cannot determine.
  • The MHD result implies that some unobserved physical parameters can be left out of the model without making prediction impossible, although accuracy drops compared with full-information forecasts.
  • The mode-truncation experiments show that small-scale structure is lost mostly because of the number of Fourier modes retained, pointing to memory-efficient architectures as the route to finer filament structure.
  • All models beat the identity baseline even though the projected snapshots appear visually similar, meaning the learned operators capture evolution beyond simple persistence.

Reading between the lines

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

  • An implicit testable consequence: if the magnetic field's orientation or strength were varied across the training set, the fixed-operator assumption could fail whenever two different 3D states share nearly the same projection; the paper's constant-B MHD setup does not probe that failure mode.
  • Because real astronomical observations are essentially static snapshots, the single-observation experiment in Appendix D is closer to the actual use case; the FNO-3D there keeps the dynamic range while the autoregressive model suppresses it, suggesting architecture choice matters more in the low-information regime.
  • The power-spectrum analysis suggests that mode count, not model capacity, is the main lever for small-scale fidelity, so progress on memory-efficient Fourier layers may directly improve filament resolution.
  • One could extend the approach to synthetic observations with radiative transfer and noise; the authors note this as future work, and it is the natural next step toward forecasting real telescope data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript applies Fourier neural operators (FNO-3D, autoregressive FNO-3D, and UNO) to predict the temporal evolution of two-dimensional projected observables—integrated density and plane-of-sky velocity—extracted from three-dimensional isothermal hydrodynamics and MHD simulations. It covers three setups: spherical gravitational collapse, supersonic turbulence, and magnetized turbulence, with the magnetic field deliberately excluded from the MHD model inputs. The paper reports normalized RMSE values of order 0.05–0.3, compares them with an identity baseline and with Gizmo solver runtimes, and concludes that neural operators can forecast observational proxies and “predict the effects of unobserved dynamical variables.”

Significance. If substantiated, the paper makes a useful contribution by moving operator learning from benchmark PDEs to high-dynamic-range astrophysical proxies and by showing that projection into observation space need not destroy short-horizon predictability in practice. The manuscript's strengths are its reproducibility (code and data links), its systematic baselines (identity mapping, power spectra), and its explicit cost-accuracy and dynamic-range ablations. The main significance claim, however, is not yet supported: the MHD experiment varies no hidden parameter, and the assumed existence of a closed operator on projected observables is asserted rather than tested.

major comments (3)
  1. [Section IV A; Table I] The text in Section IV A states that “all three models outperform the Identity mapping for all reported values,” but Table I contradicts this claim in two places. In the Spherical Collapse block, AR-FNO-3D reports δσ = 0.089, identical to the Identity baseline's δσ = 0.089; in the Turbulence block, UNO reports δmax = 0.702, which is larger than the Identity baseline's δmax = 0.622. The statement should be corrected, and the paper should discuss whether these tied or worse-than-identity metrics affect the claimed advantage of the learned surrogates.
  2. [Section II D; Section V] The conclusion in Section V that “NOs do not require complete information covering all of the physical quantities that dictate the evolution of magnetized turbulent flow” rests entirely on the MHD experiment, but Section II D fixes the initial magnetic field to B = (10^-4, 10^-4, 10^-4) G for every training series. With the hidden variable constant, the model can memorize the dynamics of a single B value without representing B at all; accuracy on held-out snapshots of the same B does not establish prediction of unobserved dynamical variables. I recommend adding experiments with varied B strengths or orientations—or at minimum evaluating on a held-out B—before the abstract's claim can be supported.
  3. [Section II A, Eq. (1)] The learning problem is set up by assuming “that there exists a nonlinear operator G : V → V that forward propagates the observations.” Because the projection φ is non-invertible, two different 3D states can have the same (σ, ū, v̄) projection while their future projected evolution diverges (for example, because of a different hidden magnetic field or line-of-sight velocity). When that occurs, no fixed operator G exists and Eq. (1) minimizes against an inconsistent target. The paper never probes this failure mode, and the constant-B design of Section II D specifically removes the only unobserved variable that could reveal it. At minimum, the authors should add a diagnostic that compares prediction errors for observationally close states with different hidden variables, or explicitly restrict the claim to fixed hidden-parameter families.
minor comments (5)
  1. [Eq. (10); Table I] The definition of δmax is not given by Eq. (10), and the phrase “the maximum error field-wise average of the test set” is ambiguous. Please clarify whether δmax is the maximum over test samples, over fields, or over timesteps; this matters because UNO's δmax in the Turbulence block exceeds the Identity baseline.
  2. [Section II B; Figure 2] Eq. (6) defines the target as [15Δt, 19Δt], but Figure 2 shows only t = 15Δt, 17Δt, and 18Δt; please state which of the five target timesteps are displayed and why.
  3. [Appendix D; Figure 8] The Figure 8 caption says the AR-FNO-3D prediction “spans a very small range of [10−4, 103]”; this interval is internally inconsistent and probably has a typo in the upper exponent.
  4. [Appendix G] The heading “Plots for MHD Turbulunce” contains a typo; it should be “Turbulence.”
  5. [Section V] The statement that “No prior works to date have investigated fluid flows with a dynamic range larger than two orders of magnitude” is an unsupported literature claim; please add a citation or qualify it with “to our knowledge.”

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the operator is learned from projected simulation pairs and evaluated on held-out data, with no fitted parameter renamed as a prediction.

full rationale

The paper's derivation chain is: run Gizmo simulations (spherical collapse, HD turbulence, MHD turbulence) -> project density and velocity to 2D proxies via Eq. (2) -> train FNO/UNO weights by minimizing the relative L2/H1 loss in Eq. (1) on (phi(f_t), phi(f_{t+1})) pairs -> report N-RMSE from Eq. (10) on a held-out split (Appendix A: 80% training, 10% validation, 10% testing, with testing data 'reserved until all training was completed'). The central claim that NOs predict the evolution of projected observables is an empirical generalization claim, not a quantity recovered from its own definition: the target phi(f_{t+1}) is not a fitted parameter, and no test label enters the optimization. The existence of a fixed operator G on the projected observation space is explicitly an assumption in Section II A, and the experiments are a test of that assumption rather than a derivation from it; even if the assumption is questionable for non-invertible projections, that is a well-posedness or experimental-design limitation, not circular reasoning. The MHD experiment fixes B = (1e-4, 1e-4, 1e-4) G, so the headline statement about 'unobserved dynamical variables' is only weakly supported, but the constant-B design does not make the reported predictions equivalent to the training labels. Self-citations (Beaumont et al. 2013, Lane et al. 2021, Xu et al. 2020-2023, Grudic et al. 2021) are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work, and no equation-level reduction of a prediction to an input can be exhibited. The honest finding is therefore no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The free parameters are model hyperparameters, not physical constants; they were tuned by random search and control the spectral truncation that drives the small-scale smoothing seen in the power spectra. The axioms are dominated by the assumed existence of G on the non-invertible projected space, which is the load-bearing assumption of the whole observational-proxy framework. No new physical entities are introduced.

free parameters (4)
  • Fourier mode truncation (FNO-3D) = 46 modes (HD), 18 modes (MHD)
    Tuned by random search; truncating high-frequency modes causes the power-spectrum falloff at k>4 documented in Figures 4 and 6, directly limiting small-scale structure.
  • Fourier mode truncation (AR-FNO-3D) = 40 modes (HD), 39 modes (MHD)
    Same as above; AR-FNO retains the most spectral slope in MHD (Figure 6).
  • Fourier mode truncation (UNO) = 17 modes (HD), 12 modes (MHD)
    Same as above; UNO's 17 modes still preserve more power than FNOs in the HD run (Figure 4), while 12 modes in MHD act like a stop-gap filter (Figure 6).
  • Network width (hidden channels) = Best at 48 (FNO-3D), 96 (AR-FNO, UNO)
    Chosen by random search / grid search in Appendix B; width controls overfitting and the cost-accuracy trade-off.
assumptions (5)
  • domain assumption There exists a fixed nonlinear operator G on the observation space such that φ(f_{t+1}) = G(φ(f_t)).
    Section II A states this assumption directly; since φ is non-invertible, the existence of an autonomous operator on projected data is not guaranteed by the underlying PDEs.
  • standard math The FNO/UNO architectures used here can approximate G sufficiently well at 64x64 resolution with the chosen hyperparameters.
    Universal approximation results for neural operators (Chen '95, Kovachki et al.) are invoked; the practical approximation quality is an empirical question.
  • domain assumption Gizmo meshless finite-mass simulations faithfully represent isothermal, self-gravitating Euler and ideal MHD dynamics.
    The training data are treated as ground truth; the paper does not validate Gizmo outputs against observations or other solvers.
  • domain assumption The simple column-density line-of-sight sum is an adequate proxy for astronomical observations.
    Section II A and Figure 1; the paper acknowledges simplification (no radiative transfer, no noise) but does not test robustness to these effects.
  • standard math An 80/10/10 split of time series gives independent train/test statistics.
    Standard supervised learning assumption; assumes no leakage between series beyond shared physics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modeling turbulent and self-gravitating fluids with Fourier neural operators." pith.science (2026). https://pith.science/paper/TKYVDUGP

@misc{pith2026250723662,
  author       = {Pith},
  title        = {Pith review of: Modeling turbulent and self-gravitating fluids with Fourier neural operators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TKYVDUGP}},
  note         = {Machine review of arXiv:2507.23662}
}
read the original abstract

Neural Operators (NOs) are a leading method for surrogate modeling of partial differential equations. Unlike traditional neural networks, which approximate individual functions, NOs learn the mappings between function spaces. While NOs have been predominantly tested on simplified 1D and 2D problems, such as those explored in prior works, these studies fail to address the complexities of more realistic, high-dimensional, and high-dynamic range systems. Moreover, many real-world applications involve incomplete or noisy data, which has not been adequately explored in current NO literature. In this work, we present a novel application of NOs to astrophysical data, which involves high-dynamic range projections into an observational space. We train Fourier NO (FNO) models to predict the evolution of incomplete observational proxies with density variations spanning four orders of magnitude. We demonstrate that FNOs can predict the effects of unobserved dynamical variables. Our work lays the groundwork for future studies that forecast direct astronomical observables.

Figures

Figures reproduced from arXiv: 2507.23662 by the authors.

Figure 1
Figure 1. FIG. 1. A schematic illustrating the approach used to generate proxy [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Cross section slices for a collapsing sphere, where time increases from left to right. The first column is the final timestep input into the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Temporal evolution of the integrated density after the system reaches a quasi-steady turbulence state. The timestep is [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Left: Power spectrum of the hydrodynamic turbulence model predictions where [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Temporal evolution of the integrated density after the magnetized fluid reaches a quasi-steady turbulence state. The timestep is [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Left: Power spectrum of the MHD turbulent models. Right: Ratio of the power predicted to the input power. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Cost accuracy trade-off for the MHD models. Fieldwise N [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Temporal evolution from left to right of the single-observation input test. The first column is the initial condition input (input five [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Temporal evolution from left to right of the first projected ve [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 12
Figure 12. Figure 12: FIG. 12. Temporal evolution from left to right of the second pro [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: FIG. 13. Temporal evolution from left to right of the first projected velocity component after the magnetized fluid reaches a quasi-steady [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: FIG. 14. Temporal evolution from left to right of the second pro [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 6 canonical work pages

  1. [1]

    merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  2. [2]

    merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  3. [3]

    merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  4. [4]

    merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  5. [5]

    Chen \ and\ author H

    author author T. Chen \ and\ author H. Chen ,\ title title Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems , \ 10.1109/72.392253 journal journal IEEE Transactions on Neural Networks \ volume 6 ,\ pages 911--917 ( year 1995 ) NoStop

    doi:10.1109/72.392253

  6. [6]

    Lu , author P

    author author L. Lu , author P. Jin , \ and\ author G. E. \ Karniadakis ,\ title title Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators , \ 10.1038/s42256-021-00302-5 journal journal Nature Machine Intelligence \ volume 3 ,\ pages 218--229 ( year 2019 ) NoStop

    doi:10.1038/s42256-021-00302-5

  7. [7]

    Kovachki , author Z

    author author N. Kovachki , author Z. Li , author B. Liu , author K. Azizzadenesheli , author K. Bhattacharya , author A. Stuart , \ and\ author A. Anandkumar ,\ title title Neural operator: Learning maps between function spaces with applications to pdes , \ http://jmlr.org/papers/v24/21-1524.html journal journal Journal of Machine Learning Research \ vol...

    work page 2023

  8. [8]

    Li , author N

    author author Z. Li , author N. Kovachki , author K. Azizzadenesheli , author B. Liu , author K. Bhattacharya , author A. Stuart , \ and\ author A. Anandkumar ,\ @noop title Fourier neural operator for parametric partial differential equations , \ ( year 2021 ),\ http://arxiv.org/abs/2010.08895 arXiv:2010.08895 [cs.LG] NoStop

    arXiv 2021

Show all 48 references
  1. [9]

    O'Leary-Roseberry , author P

    author author T. O'Leary-Roseberry , author P. Chen , author U. Villa , \ and\ author O. Ghattas ,\ @noop title Derivative-informed neural operator: An efficient framework for high-dimensional parametric derivative learning , \ ( year 2023 ),\ http://arxiv.org/abs/2206.10745 a...

    2023 arXiv

  2. [10]

    Takamoto , author T

    author author M. Takamoto , author T. Praditia , author R. Leiteritz , author D. MacKinlay , author F. Alesiani , author D. Pflüger , \ and\ author M. Niepert ,\ @noop title Pdebench: An extensive benchmark for scientific machine learning , \ ( year 2023 ),\ http://arxiv.org/a...

    2023 arXiv

  3. [11]

    Bhattacharya , author B

    author author K. Bhattacharya , author B. Hosseini , author N. B. \ Kovachki , \ and\ author A. M. \ Stuart ,\ @noop title Model reduction and neural networks for parametric pdes , \ ( year 2021 ),\ http://arxiv.org/abs/2005.03180 arXiv:2005.03180 [math.NA] NoStop

    2021 arXiv

  4. [12]

    author author M. V. \ de Hoop , author D. Z. \ Huang , author E. Qian , \ and\ author A. M. \ Stuart ,\ @noop title The cost-accuracy trade-off in operator learning with neural networks , \ ( year 2022 ),\ http://arxiv.org/abs/2203.13181 arXiv:2203.13181 [math.NA] NoStop

    2022 arXiv

  5. [14]

    Li , author M

    author author Z. Li , author M. Liu-Schiaffini , author N. Kovachki , author B. Liu , author K. Azizzadenesheli , author K. Bhattacharya , author A. Stuart , \ and\ author A. Anandkumar ,\ @noop title Learning dissipative dynamics in chaotic systems , \ ( year 2022 ),\ http://...

    2022 arXiv

  6. [15]

    Gopakumar , author S

    author author V. Gopakumar , author S. Pamela , author L. Zanisi , author Z. Li , author A. Anandkumar , \ and\ author M. Team ,\ @noop title Fourier neural operator for plasma modelling , \ ( year 2023 a ),\ http://arxiv.org/abs/2302.06542 arXiv:2302.06542 [physics.plasm-ph] NoStop

    2023 arXiv

  7. [16]

    Gopakumar , author S

    author author V. Gopakumar , author S. Pamela , author L. Zanisi , author Z. Li , author A. Gray , author D. Brennand , author N. Bhatia , author G. Stathopoulos , author M. Kusner , author M. P. \ Deisenroth , author A. Anandkumar , author J. Team , \ and\ author M. Team ,\ @...

    2023 arXiv

  8. [17]

    author author S. G. \ Rosofsky \ and\ author E. A. \ Huerta ,\ title title Magnetohydrodynamics with physics informed neural operators , \ 10.1088/2632-2153/ace30a journal journal Machine Learning: Science and Technology \ volume 4 ,\ pages 035002 ( year 2023 ) NoStop

    doi

  9. [18]

    Wen , author Z

    author author G. Wen , author Z. Li , author K. Azizzadenesheli , author A. Anandkumar , \ and\ author S. M. \ Benson ,\ @noop title U-fno -- an enhanced fourier neural operator-based deep-learning model for multiphase flow , \ ( year 2022 ),\ http://arxiv.org/abs/2109.03697 a...

    2022 arXiv

  10. [19]

    Bonev , author T

    author author B. Bonev , author T. Kurth , author C. Hundt , author J. Pathak , author M. Baust , author K. Kashinath , \ and\ author A. Anandkumar ,\ @noop title Spherical fourier neural operators: Learning stable dynamics on the sphere , \ ( year 2023 ),\ http://arxiv.org/ab...

    2023 arXiv

  11. [20]

    Brandstetter , author R

    author author J. Brandstetter , author R. van den Berg , author M. Welling , \ and\ author J. K. \ Gupta ,\ @noop title Clifford neural layers for pde modeling , \ ( year 2023 ),\ http://arxiv.org/abs/2209.04934 arXiv:2209.04934 [cs.LG] NoStop

    2023 arXiv

  12. [21]

    Pathak , author S

    author author J. Pathak , author S. Subramanian , author P. Harrington , author S. Raja , author A. Chattopadhyay , author M. Mardani , author T. Kurth , author D. Hall , author Z. Li , author K. Azizzadenesheli , author P. Hassanzadeh , author K. Kashinath , \ and\ author A. ...

    2022 arXiv

  13. [22]

    Zhao , author Z

    author author S. Zhao , author Z. Li , author B. Fan , author Y. Wang , author H. Yang , \ and\ author J. Wang ,\ https://arxiv.org/abs/2411.04502 title Lesnets (large-eddy simulation nets): Physics-informed neural operator for large-eddy simulation of turbulence , \ ( year 20...

    2025 arXiv

  14. [23]

    Wang , author J

    author author C. Wang , author J. Berner , author Z. Li , author D. Zhou , author J. Wang , author J. Bae , \ and\ author A. Anandkumar ,\ https://arxiv.org/abs/2408.05177 title Beyond closure models: Learning chaotic-systems via physics-informed neural operators , \ ( year 20...

    2024 arXiv

  15. [24]

    Molinaro , author S

    author author R. Molinaro , author S. Lanthaler , author B. Raonić , author T. Rohner , author V. Armegioiu , author S. Simonis , author D. Grund , author Y. Ramic , author Z. Y. \ Wan , author F. Sha , author S. Mishra , \ and\ author L. Zepeda-Núñez ,\ https://arxiv.org/abs/...

    2025 arXiv

  16. [25]

    author author C. F. \ McKee \ and\ author E. C. \ Ostriker ,\ title title Theory of star formation , \ 10.1146/annurev.astro.45.051806.110602 journal journal Annual Review of Astronomy and Astrophysics \ volume 45 ,\ pages 565–687 ( year 2007 ) NoStop

    arXiv

  17. [26]

    author author P. F. \ Hopkins \ and\ author M. J. \ Raives ,\ title title Accurate, meshless methods for magnetohydrodynamics , \ 10.1093/mnras/stv2180 journal journal Monthly Notices of the Royal Astronomical Society \ volume 455 ,\ pages 51–88 ( year 2015 ) NoStop

    doi

  18. [27]

    author author J. R. \ Beattie , author C. Federrath , author R. S. \ Klessen , author S. Cielo , \ and\ author A. Bhattacharjee ,\ https://arxiv.org/abs/2405.16626 title Magnetized compressible turbulence with a fluctuation dynamo and reynolds numbers over a million , \ ( year...

    2024 arXiv

  19. [28]

    author author M. Y. \ Grudić , author D. Guszejnov , author P. F. \ Hopkins , author S. S. R. \ Offner , \ and\ author C.-A. \ Faucher-Giguère ,\ title title Starforge: Towards a comprehensive numerical model of star cluster formation and feedback , \ 10.1093/mnras/stab1347 jo...

    doi

  20. [29]

    author author P. F. \ Hopkins , author J. Squire , author K.-Y. \ Su , author U. P. \ Steinwandel , author K. Kremer , author Y. Shi , author M. Y. \ Grudic , author S. Wellons , author C.-A. \ Faucher-Giguere , author D. Angles-Alcazar , author N. Murray , \ and\ author E. Qu...

  21. [30]

    author author E. J. \ Weber \ and\ author L. J. \ Davis ,\ title title The Angular Momentum of the Solar Wind , \ 10.1086/149138 journal journal The Astrophysical Journal \ volume 148 ,\ pages 217--227 ( year 1967 ) NoStop

    1967 doi

  22. [31]

    M.Stone , author K

    author author J. M.Stone , author K. Tomida , author C. J. \ White , \ and\ author K. G. \ Felker ,\ title title The athena++ adaptive mesh refinement framework: Design and magnetohydrodynamic solvers , \ 10.3847/1538-4365/ab929b journal journal The Ast. J. \ volume 249 ,\ eid...

    arXiv

  23. [32]

    Fryxell , author K

    author author B. Fryxell , author K. Olson , author P. Ricker , author F. X. \ Timmes , author M. Zingale , author D. Q. \ Lamb , author P. MacNeice , author R. Rosner , author J. W. \ Truran , \ and\ author H. Tufo ,\ title title Flash: An adaptive mesh hydrodynamics code for...

    doi

  24. [33]

    author author P. F. \ Hopkins ,\ title title A new class of accurate, mesh-free hydrodynamic simulation methods , \ 10.1093/mnras/stv195 journal journal Monthly Notices of the Royal Astronomical Society \ volume 450 ,\ pages 53–110 ( year 2015 ) NoStop

    doi

  25. [34]

    author author C. N. \ Beaumont , author S. S. R. \ Offner , author R. Shetty , author S. C. O. \ Glover , \ and\ author A. A. \ Goodman ,\ title title Quantifying observational projection effects using molecular cloud simulations , \ 10.1088/0004-637X/777/2/173 journal journal...

    doi

  26. [35]

    author author T. J. \ Haworth , author S. C. \ Glover , author C. M. \ Koepferl , author T. G. \ Bisbas , \ and\ author J. E. \ Dale ,\ title title Synthetic observations of star formation and the interstellar medium , \ 10.1016/j.newar.2018.06.001 journal journal New Astronom...

    doi

  27. [36]

    Xu , author S

    author author D. Xu , author S. S. R. \ Offner , author R. Gutermuth , \ and\ author C. V. \ Oort ,\ title title Application of convolutional neural networks to identify stellar feedback bubbles in co emission , \ 10.3847/1538-4357/ab6607 journal journal The Astrophysical Jour...

    doi

  28. [37]

    Xu , author S

    author author D. Xu , author S. S. R. \ Offner , author R. Gutermuth , \ and\ author C. V. \ Oort ,\ title title Application of convolutional neural networks to identify protostellar outflows in co emission , \ 10.3847/1538-4357/abc7bf journal journal The Astrophysical Journal...

    doi

  29. [38]

    Xu , author S

    author author D. Xu , author S. Offner , author R. Gutermuth , author M. Grudic , author D. Guszejnov , \ and\ author P. Hopkins ,\ @noop title Predicting the radiation field of molecular clouds using denoising diffusion probabilistic models , \ ( year 2023 a ),\ http://arxiv....

    2023 arXiv

  30. [39]

    Xu , author C.-Y

    author author D. Xu , author C.-Y. \ Law , \ and\ author J. C. \ Tan ,\ title title Application of convolutional neural networks to predict magnetic fields’ directions in turbulent clouds , \ 10.3847/1538-4357/aca66c journal journal The Astrophysical Journal \ volume 942 ,\ pa...

    doi

  31. [40]

    Xu , author J

    author author D. Xu , author J. C. \ Tan , author C.-J. \ Hsu , \ and\ author Y. Zhu ,\ title title Denoising diffusion probabilistic models to predict the density of molecular clouds , \ 10.3847/1538-4357/accae5 journal journal The Astrophysical Journal \ volume 950 ,\ pages ...

    doi

  32. [41]

    Hersbach , author B

    author author H. Hersbach , author B. Bell , author P. Berrisford , author S. Hirahara , author A. Horányi , author J. Muñoz-Sabater , author J. Nicolas , author C. Peubey , author R. Radu , author D. Schepers , author A. Simmons , author C. Soci , author S. Abdalla , author X...

  33. [42]

    author author M. J. \ Turk , author B. D. \ Smith , author J. S. \ Oishi , author S. Skory , author S. W. \ Skillman , author T. Abel , \ and\ author M. L. \ Norman ,\ title title yt: A multi-code analysis toolkit for astrophysical simulation data , \ 10.1088/0067-0049/192/1/9...

    doi

  34. [43]

    author author H. B. \ Lane , author M. Y. \ Grudić , author D. Guszejnov , author S. S. R. \ Offner , author C.-A. \ Faucher-Giguère , \ and\ author A. L. \ Rosen ,\ title title Less wrong: a more realistic initial condition for simulations of turbulent molecular clouds , \ 10...

    doi

  35. [44]

    Kossaifi , author N

    author author J. Kossaifi , author N. Kovachki , author K. Azizzadenesheli , \ and\ author A. Anandkumar ,\ @noop title Multi-grid tensorized fourier neural operator for high-resolution pdes , \ ( year 2023 ),\ http://arxiv.org/abs/2310.00120 arXiv:2310.00120 [cs.LG] NoStop

    2023 arXiv

  36. [45]

    author author M. A. \ Rahman , author Z. E. \ Ross , \ and\ author K. Azizzadenesheli ,\ @noop title U-no: U-shaped neural operators , \ ( year 2023 ),\ http://arxiv.org/abs/2204.11127 arXiv:2204.11127 [cs.LG] NoStop

    2023 arXiv

  37. [46]

    White , author R

    author author C. White , author R. Tu , author J. Kossaifi , author G. Pekhimenko , author K. Azizzadenesheli , \ and\ author A. Anandkumar ,\ @noop title Speeding up fourier neural operators via mixed precision , \ ( year 2023 ),\ http://arxiv.org/abs/2307.15034 arXiv:2307.15...

    2023 arXiv

  38. [47]

    Liu-Schiaffini , author J

    author author M. Liu-Schiaffini , author J. Berner , author B. Bonev , author T. Kurth , author K. Azizzadenesheli , \ and\ author A. Anandkumar ,\ https://arxiv.org/abs/2402.16845 title Neural operators with localized integral and differential kernels , \ ( year 2024 ),\ http...

    2024 arXiv

  39. [48]

    author author D. P. \ Kingma \ and\ author J. Ba ,\ @noop title Adam: A method for stochastic optimization , \ ( year 2017 ),\ http://arxiv.org/abs/1412.6980 arXiv:1412.6980 [cs.LG] NoStop

    2017 arXiv

  40. [49]

    Loshchilov \ and\ author F

    author author I. Loshchilov \ and\ author F. Hutter ,\ @noop title Sgdr: Stochastic gradient descent with warm restarts , \ ( year 2017 ),\ http://arxiv.org/abs/1608.03983 arXiv:1608.03983 [cs.LG] NoStop

    2017 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.