Pith. sign in

REVIEW 3 major objections 4 minor 40 references

A 3D convolutional network predicts solar subgrid Reynolds stresses more accurately than standard physics-based closures.

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 · grok-4.5

2026-07-13 05:24 UTC pith:DWJ6QZAZ

load-bearing objection Solid offline a-priori CNN gains on quiet-Sun Reynolds stresses, but the live-closure claim is still untested. the 3 major comments →

arxiv 2607.08969 v1 pith:DWJ6QZAZ submitted 2026-07-09 astro-ph.SR physics.flu-dyn

Developing Machine Learning Models of Subgrid Turbulent Transport for Quiet Sun 3D Radiative Hydrodynamic Simulations

Rimsha Hameed Syeda , Dustin Kempton , Viacheslav Sadykov , Irina Kitiashvili , Rafal Angryk This is my paper
classification astro-ph.SR physics.flu-dyn
keywords subgrid-scale turbulenceReynolds stress tensor3D convolutional neural networksquiet Sunradiative hydrodynamicsLarge Eddy SimulationSmagorinsky model
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.

Solar plasma simulations cannot resolve every scale of turbulent motion, so they need a cheap estimate of the missing small-scale stresses that still transport momentum and dissipate energy. This paper shows that a carefully designed three-dimensional convolutional neural network can learn those stresses directly from high-resolution quiet-Sun radiative-hydrodynamic runs and then reconstruct the six components of the Reynolds stress tensor more faithfully than either a classical Gradient model or the Smagorinsky eddy-viscosity model. The network is trained on local 3-by-3-by-3 velocity neighborhoods plus density, after a signed-log transform that tames the heavily skewed stress distributions. On held-out data the best architecture cuts root-mean-square error by roughly thirty percent on the diagonal stresses and eight percent on the off-diagonal stresses relative to the Gradient baseline. The result matters because it opens a practical route to cheaper yet more faithful large-eddy simulations of the solar convection zone and lower atmosphere.

Core claim

A 3D convolutional neural network that ingests local averaged velocity cubes and density can reconstruct the six Reynolds stress tensor components of quiet-Sun turbulence more accurately than the Gradient model, the Smagorinsky model (under two coefficient choices), and a multilayer perceptron, delivering average RMSE reductions of about 31 percent on the diagonal components and 8 percent on the off-diagonal components.

What carries the argument

The 3DCNN1.1 architecture: two 3-by-3-by-3 convolutional layers (32 then 64 filters, LeakyReLU) that extract spatial features from the three velocity channels, followed by concatenation of the central density and SoftSign fully-connected layers that output the six stress components; trained after signed-log transformation of the targets.

Load-bearing premise

That accurate offline predictions on filtered high-resolution cubes (and static low-resolution snapshots) will remain accurate once the same network is inserted as a live subgrid closure inside a time-evolving solar simulation.

What would settle it

Insert the trained network as the active subgrid model inside a low-resolution StellarBox run and check whether the resulting large-scale velocity and density statistics, plus the emergent stress distributions, match those of a high-resolution reference simulation better than the same run performed with a Smagorinsky closure.

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

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 / 4 minor

Summary. The paper develops 3D convolutional neural network (3DCNN) surrogates, along with an MLP baseline, to predict the six components of the Reynolds stress tensor τ_ij from local 3×3×3 averaged velocity fields and central density in quiet-Sun StellarBox radiative hydrodynamic simulations. High-resolution (∼12.5 km) cubes are filtered to ∼50 km effective resolution to generate both inputs and exact targets via the definition τ_ij = 〈u_i u_j〉 − ũ_i ũ_j. After log/signed-log transforms and standardization of the heavily skewed targets, the best architecture (3DCNN1.1 with LeakyReLU) is shown via held-out RMSE, R², PDFs, error histograms, density-binned diagnostics, and low-resolution snapshot tests to outperform the Gradient model, Smagorinsky (Cs=Cc=0.1 and 0.001), and MLP, with average RMSE reductions of ∼31 % on diagonal and ∼8 % on off-diagonal components relative to Gradient (Table 5). The authors conclude that CNNs are a viable candidate for subgrid-scale modeling, while explicitly deferring online coupling into StellarBox.

Significance. If the offline gains translate under live coupling, the work would supply a coefficient-free, data-driven alternative to classical Smagorinsky/Gradient closures for solar convection-zone LES, with clear practical value for multi-scale quiet-Sun modeling. Even as a pure a-priori study the manuscript is carefully executed: physical units are restored after inverted transforms, multiple physics baselines and architecture ablations are reported, and density-binned plus cluster diagnostics expose regime-dependent performance. These elements constitute a solid, reproducible foundation for subsequent a-posteriori tests and are already useful to the solar and LES communities.

major comments (3)
  1. [Abstract, §1, §5.4] Abstract, §1 and §5.4: the central claim that the 3DCNN is “a viable candidate for modeling subgrid processes and a promising alternative to traditional turbulence models” rests exclusively on a-priori regression accuracy (Table 5, Figs. 9–12). When the same network is applied to genuine low-resolution (∼50 km) fields the predicted stress PDFs systematically compress in the tails (Figs. 13–14). Classical LES literature treats the a-priori/a-posteriori gap as decisive; without at least one online StellarBox integration (or a clear, quantitative statement that the present results do not yet establish live-closure viability) the abstract and conclusion over-reach the evidence that is actually supplied.
  2. [Table 5, §5.1] Table 5 and §5.1: the headline ∼31 % / ∼8 % RMSE reductions are computed solely against the Gradient model. Against Smagorinsky (Cs=Cc=0.001) the gains are larger, yet that coefficient set is known to under-dissipate; against Cs=Cc=0.1 the diagonal gains shrink. The abstract and conclusion should report the full range of relative improvements (or at least both Smagorinsky settings) so that the claimed superiority is not tied to a single, relatively weak baseline.
  3. [§3.1, Eq. (5)] §3.1 and Eq. (5): targets are obtained by direct spatial averaging of the high-resolution velocity products. While this is the standard a-priori procedure, the manuscript never quantifies how sensitive the learned mapping is to the precise filter kernel or to the non-overlapping 4×4×4 sub-cube sampling. A short sensitivity test (or an explicit statement that the reported RMSE is filter-specific) is needed before the numbers can be treated as robust estimates of subgrid stress.
minor comments (4)
  1. [Figs. 1–2] Figure 1 and Figure 2 captions: “normalized units” are defined only later in the text; a one-sentence reminder in the captions would improve readability.
  2. [§5.7] §5.7: the K-means analysis is performed on a single data cube and uses an arbitrary k=5; the section title already labels it “preliminary,” but the body still presents cluster-wise R² values as if they generalize. Soften the language or move the entire subsection to an appendix.
  3. [Table 6] Table 6: RMSLE is introduced without definition; a brief formula or reference would help non-ML readers.
  4. Throughout: occasional typographical inconsistencies (e.g., “Th eAstrophysical Journal”, missing spaces around ×10^n) should be cleaned in production.

Circularity Check

0 steps flagged

No circularity: supervised regression of independently defined Reynolds stresses against held-out data and external physics baselines.

full rationale

The paper computes Reynolds-stress targets directly from high-resolution velocity products via the standard definition (Eq. 5), trains a 3DCNN (and MLP) to regress those targets from filtered 3 imes3 imes3 velocity+density inputs, and evaluates RMSE/R^{2} on a held-out test set after inverting the log transform. The reported ~31 % / ~8 % improvements (Table 5, abstract) are empirical comparisons against the Gradient model (Eq. 6) and two fixed-coefficient Smagorinsky models (Eqs. 7–9); no free parameter is fitted to force agreement with the claimed gains, and the log/signed-log preprocessing is a conventional variance-stabilizing step whose effect is ablated (Sec. 5.6, Table 6). Self-citations supply only the StellarBox data source and prior solar-simulation context; they do not underwrite uniqueness theorems, ansatzes, or load-bearing mathematical steps. The a-priori / a-posteriori gap noted by the skeptic is a validity limitation of the experimental design, not a circular reduction of the derivation chain. The central claim therefore stands as ordinary supervised learning plus external benchmarking and contains no self-definitional, fitted-input-as-prediction, or self-citation-load-bearing circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central performance claim rests on standard LES filtering assumptions, the classical definition of Reynolds stress, two hand-chosen Smagorinsky coefficient pairs used only as baselines, and a suite of architectural and preprocessing choices that were tuned on the same data family. No new physical entities are postulated; the free parameters are ordinary ML hyper-parameters and the two baseline coefficients.

free parameters (4)
  • Smagorinsky coefficients Cs=Cc = 0.1 and 0.001
    Two discrete pairs (0.1 and 0.001) are chosen by hand as physics baselines; the ML claim is relative to these choices.
  • CNN learning rate, batch size, filter counts, activation slopes = lr=0.001, batch=128, alpha=0.01
    Tuned via grid/early-stopping on the training set; final values (lr=0.001, batch=128, 32/64 filters, LeakyReLU alpha=0.01) directly affect reported MSE/R2.
  • Log / signed-log offset and scaling constants for targets = scale 1e8, offset +1
    Division by 1e8 and +1 offset for off-diagonal terms are chosen to stabilize training; they alter the loss landscape and final inverted metrics.
  • K-means cluster count (k=5) = 5
    Used only for post-hoc analysis; still a free choice that partitions the reported per-regime errors.
axioms (4)
  • domain assumption Reynolds stress is exactly tau_ij = <u_i u_j> - <u_i><u_j> computed on the high-resolution cubes after 4x spatial averaging.
    Section 2.2 and Eq. 5; the entire supervised target is defined by this filtering operation.
  • domain assumption Non-overlapping 4x4x4 sub-cubes sampled every 5 min are statistically independent enough for random train/val/test splits.
    Section 3.1; underpins the claim that held-out metrics generalize.
  • ad hoc to paper Offline regression accuracy on filtered fields is a meaningful indicator of utility as an SGS closure.
    Stated as the evaluation protocol while full online coupling is deferred (Sections 1, 5.4).
  • domain assumption Standard LES continuum equations and the compressible Smagorinsky form used by StellarBox remain valid for the quiet-Sun regime studied.
    Background Sections 2.1-2.2; inherited from prior StellarBox literature.

pith-pipeline@v1.1.0-grok45 · 27112 in / 2932 out tokens · 43471 ms · 2026-07-13T05:24:28.742335+00:00 · methodology

0 comments
read the original abstract

Numerical modeling of solar plasma dynamics is affected by the resolution of the computational grid. This often requires the estimation of subgrid processes related to the small-scale flow turbulence, as these processes play a critical role in momentum transport and energy dissipation. In this work, we investigate the use of deep learning techniques as surrogate models for subgrid turbulent transport in realistic hydrodynamic simulations of the quiet Sun. We describe the development of a 3D Convolutional Neural Network (CNN) to capture spatial dependencies in 3D velocity fields, leveraging different activation functions, as well as different architectural designs. We specifically focus on the prediction of Reynolds stress tensor components. The resultant model integrates velocity vector components and scalar features, such as plasma density, to enhance prediction accuracy. We compare the 3DCNN model to other types of models, such as a Multilayer Perceptron (MLP) and physics-based Gradient and Smagorinsky models, and show that the final model design reconstructs the Reynolds stress tensor components more accurately. Specifically, a 3DCNN model achieves an average improvement of ~31% on diagonal components and ~8% on the off-diagonal components of the stress tensor. Additionally, we show that applying a logarithmic data transformation of the target stress tensor components, to handle heavily skewed data, improves model performance. Results demonstrate the potential of deep learning, particularly CNNs, to approximate Reynolds stress tensor components for the upper solar convection zone and lower atmosphere, making them a viable candidate for modeling subgrid processes and a promising alternative to traditional turbulence models.

Figures

Figures reproduced from arXiv: 2607.08969 by Dustin Kempton, Irina Kitiashvili, Rafal Angryk, Rimsha Hameed Syeda, Viacheslav Sadykov.

Figure 1
Figure 1. Figure 1: Panels (a) and (b) show the original distribution of τvv and τuw. All distributions are log-scaled on y-axis for better visualization. The preprocessing stage is crucial for preparing the input and target variables extracted from the dataset for effective utilization in machine learning models. We begin by applying a Z-score normalization to all the velocity inputs (27 sub-cubes × 3 components) and density… view at source ↗
Figure 2
Figure 2. Figure 2: Panels (a) and (b) present the transformed distributions of τvv and τuw after preprocessing, respectively. Here, normalized units denote that the stresses have been log or signed-log transformed and then standardized. There was additional non-dimensionalization. 1. Logarithmic Transformation: In cases of highly skewed data with long tails, regression anal￾ysis is easier to perform using the logarithmic tra… view at source ↗
Figure 3
Figure 3. Figure 3: Distributions of Central Velocity and Density Fields when Smagorinsky coefficients are CS = CC = 0.001. settings introduced in Section 2: CS = CC = 0.1 and CS = CC = 0.001. For each case, we examine the distributions of the averaged velocity components (u, v, w) and the central cell density ρ [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distributions of Central Velocity and Density Fields when Smagorinsky coefficients are CS = CC = 0.1. velocities in the tails. This is expected: averaging over a coarser grid smooths out small-scale velocity fluctuations. The density distributions remain closely matched between resolutions. Comparing the two low-resolution runs against each other in [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of velocity and density distributions for the two low-resolution simulation datasets. 3.3. Model Description 3.3.1. Multilayer Perceptron As a simplest-case ML model, we utilize the MLPRegressor of scikit-learn (Pedregosa et al. 2011), a neural network designed for regression tasks. The Multilayer Perceptron (MLP) is a fundamental model in neural networks, consisting of multiple layers of fully … view at source ↗
Figure 6
Figure 6. Figure 6: Model schematics of 3DCNN1. In the development of 3DCNN1, multiple activation functions and layer configurations were explored to optimize the model’s ability to capture spatial hierarchies from local 3D velocity fields. In all three 3DCNN1 variants, the velocity components are first processed through two 3D convolutional layers; the first convolutional layer uses 32 filters equipped with a kernel 3 × 3 × … view at source ↗
Figure 7
Figure 7. Figure 7: Model schematics of 3DCNN2. 4. EXPERIMENTAL SETTINGS 4.1. Parameter Tuning MLP To find the best-performing MLP implementation, we performed hyperparameter tuning utilizing the built-in GridSearchCV method of scikit-learn (Pedregosa et al. 2011), which is a technique used to systematically search through a predefined grid of hyperparameters and select the combination that yields the best performance. The pa… view at source ↗
Figure 8
Figure 8. Figure 8: Final CNN architecture based on the model schematics of 3DCNN1 in [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distributions of components τvv shown in Panel(a) and τuw in (b). Blue curves show the original target data, orange shows the predictions of the MLPRegressor, and gray shows the predictions of the CNN. average MSE, lowest RMSE, and the highest R2 score) for subsequent comparative analyses with the baselines, while we should always keep in mind the related drawbacks of this selection. 5.3. Comparison of Err… view at source ↗
Figure 10
Figure 10. Figure 10: The Frequency plots of the deviations of τvv in Panel(a) and τuw in (b) between the predictions of the best CNN model (purple), Smagorinsky CS = CC = 0.1 (green), and Gradient models (red) and the target value. The Smagorinsky model (with CS = CC = 0.1) is approximately unbiased, i.e., its distribution is centered near zero, but displays the broadest distributions and heavier tails, indicating overall les… view at source ↗
Figure 11
Figure 11. Figure 11: The Frequency plots of the deviations of τvv in Panel(a) and τuw in (b) between the predictions of the best CNN model (purple), Smagorinsky CS = CC = 0.001 (green), and Gradient models (red) and the target value. 5.3.1. Scatterplots of Gradient and CNN Models Furthermore, we demonstrate the scatterplots in [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The scatter plots of two components τvv in Panel(a) and τuw in (b) comparing the predictions of the best CNN model and the Gradient model. Blue points show the predictions of CNN, and red points show the predictions of the Gradient model. predictions on τuv [PITH_FULL_IMAGE:figures/full_fig_p029_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Distributions of components τvv shown in Panel (a) and τuw in Panel (b). Blue curves show the original target distributions, orange curves show CNN predictions using filtered/coarse-grained inputs derived from the high-resolution simulations, and gray curves show CNN predictions using inputs from the genuine low-resolution simulations for the case CS = CC = 0.001. resolution data. The two low-resolution r… view at source ↗
Figure 14
Figure 14. Figure 14: Distributions of components τvv shown in Panel (a) and τuw in Panel (b). Blue curves show the original target distributions, orange curves show CNN predictions using filtered/coarse-grained inputs derived from the high-resolution simulations, and gray curves show CNN predictions using inputs from the genuine low-resolution simulations for the case CS = CC = 0.1. low-resolution input fields. To further val… view at source ↗
Figure 15
Figure 15. Figure 15: The box-plots τvv across different density bins. larger convection scales. In this subsection, we will investigate the impact that the inclusion of the densities have on the Reynolds stress tensor modeling. The box-and-whisker plots of τvv and τuw components across different density bins are shown in Figures 15 and 16. The plots reveal key trends in the behavior of turbulent stresses within the solar diff… view at source ↗
Figure 16
Figure 16. Figure 16: The box-plots τuw across different density bins. where the convection motions are already suppressed. Conversely, higher-density bins tend to show increased variability, as demonstrated by wider interquartile ranges, suggesting more heterogeneous stress distributions in these regimes (besides the last bin, which likely corresponds to the depths of 2 Mm and below, where the convective motions experience la… view at source ↗
Figure 17
Figure 17. Figure 17: The error box-plots τuu across different density bins for all models(case Smagorinsky Cs=Cc=0.1) [PITH_FULL_IMAGE:figures/full_fig_p035_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: The scatter plots of two components τvv in Panel(a) and τuw in (b) comparing the predictions of the 3DCNN1 trained with data before and after normalization. Blue dots represent the predictions of CNN when trained with normalized data, and green represents the predictions of CNN when trained without a log transform. We have investigated the performance of the selected CNN model trained using logarithmic tr… view at source ↗

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

40 extracted references · 16 canonical work pages

  1. [1]

    V., & Carlsson, M

    Gudiksen, B. V., & Carlsson, M. 2023, Astronomy & Astrophysics, 675, A103, doi: 10.1051/0004-6361/202346765 Bell´ ego, C., Benatia, D., & Pape, L. 2022, Dealing with Logs and Zeros in Regression Models, doi: 10.48550/arXiv.2203.11820 43

    work page doi:10.1051/0004-6361/202346765 2023

  2. [2]

    2010, Astronomische Nachrichten, 331, 636, doi: 10.1002/asna.201011390

    Cao, W., Gorceix, N., Coulter, R., et al. 2010, Astronomische Nachrichten, 331, 636, doi: 10.1002/asna.201011390

    doi:10.1002/asna.201011390 2010

  3. [3]

    J., & Zwinderman, A

    Cleophas, T. J., & Zwinderman, A. H. 2016, Logarithmic Transformations, a Great Help to Statistical Analyses (Cham: Springer International Publishing), 243–247, doi: 10.1007/978-3-319-27104-0 43

    work page doi:10.1007/978-3-319-27104-0 2016

  4. [4]

    J., & Altman, D

    Cole, T. J., & Altman, D. G. 2017, BMJ, 358, doi: 10.1136/bmj.j3683

    work page doi:10.1136/bmj.j3683 2017

  5. [5]

    Germano, M., Piomelli, U., Moin, P., & Cabot, W. H. 1991, Physics of Fluids A, 3, 1760, doi: 10.1063/1.857955

    doi:10.1063/1.857955 1991

  6. [6]

    V., Carlsson, M., Hansteen, V

    Gudiksen, B. V., Carlsson, M., Hansteen, V. H., et al. 2011, Astronomy & Astrophysics, 531, A154, doi: 10.1051/0004-6361/201116520

    doi:10.1051/0004-6361/201116520 2011

  7. [7]

    R., Millman, K

    Harris, C. R., Millman, K. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    doi:10.1038/s41586-020-2649-2 2020

  8. [8]

    2015, in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 1026–1034, doi: 10.1109/ICCV.2015.123

    He, K., Zhang, X., Ren, S., & Sun, J. 2015, in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 1026–1034, doi: 10.1109/ICCV.2015.123

    doi:10.1109/iccv.2015.123 2015

  9. [9]

    2016, in Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778, doi: 10.1109/CVPR.2016.90

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778, doi: 10.1109/CVPR.2016.90

    doi:10.1109/cvpr.2016.90 2016

  10. [10]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

    doi:10.1109/mcse.2007.55 2007

  11. [11]

    Mansour, N. N. 2008, The Astrophysical Journal Letters, 684, L51, doi: 10.1086/592042

    work page doi:10.1086/592042 2008

  12. [12]

    I., Huang, C., Sitdikov, I., et al

    Karpov, P. I., Huang, C., Sitdikov, I., et al. 2022, The Astrophysical Journal, 940, 26, doi: 10.3847/1538-4357/ac88cc

    doi:10.3847/1538-4357/ac88cc 2022

  13. [13]

    2022, Monthly Weather Review, 150, 2279 , doi: 10.1175/MWR-D-21-0217.1

    Khani, S., & Port´ e-Agel, F. 2022, Monthly Weather Review, 150, 2279 , doi: 10.1175/MWR-D-21-0217.1

    work page doi:10.1175/mwr-d-21-0217.1 2022

  14. [14]

    Kitiashvili, I. N. 2014, Publications of the Astronomical Society of Japan, 66, S8, doi: 10.1093/pasj/psu114

    work page doi:10.1093/pasj/psu114 2014

  15. [15]

    N., Abramenko, V

    Kitiashvili, I. N., Abramenko, V. I., Goode, P. R., et al. 2013, Physica Scripta Volume T, 155, 014025, doi: 10.1088/0031-8949/2013/T155/014025

    work page doi:10.1088/0031-8949/2013/t155/014025 2013

  16. [16]

    N., Couvidat, S., & Lagg, A

    Kitiashvili, I. N., Couvidat, S., & Lagg, A. 2015, Th eAstrophysical Journal, 808, 59, doi: 10.1088/0004-637X/808/1/59

    work page doi:10.1088/0004-637x/808/1/59 2015

  17. [17]

    N., Kosovichev, A

    Kitiashvili, I. N., Kosovichev, A. G., Mansour, N. N., & Wray, A. A. 2011, Solar Physics, 268, 283, doi: 10.1007/s11207-010-9679-0 —. 2012, The Astrophysical Journal Letters, 751, L21, doi: 10.1088/2041-8205/751/1/L21

    work page doi:10.1007/s11207-010-9679-0 2011

  18. [18]

    M., & Guerrero, G

    Sadykov, V. M., & Guerrero, G. 2023, MNRAS, 518, 504, doi: 10.1093/mnras/stac2946

    work page doi:10.1093/mnras/stac2946 2023

  19. [19]

    L., Hannun, A

    Maas, A. L., Hannun, A. Y., & Ng, A. Y. 2013, in Proceedings of the 30th International Conference on Machine Learning (ICML), PMLR, 3–8. http://proceedings.mlr.press/v28/maas13.html

    2013

  20. [20]

    S., & Cabot, W

    Meneveau, C., Lund, T. S., & Cabot, W. H. 1994, Journal of Fluid Mechanics, 319, 353 . https: //api.semanticscholar.org/CorpusID:122183534 44

    1994

  21. [21]

    2022, Astronomy & Astrophysics, 663, A96, doi: 10.1051/0004-6361/202243439

    Vitas, N. 2022, Astronomy & Astrophysics, 663, A96, doi: 10.1051/0004-6361/202243439

    work page doi:10.1051/0004-6361/202243439 2022

  22. [22]

    1985, Solar Physics, 100, 209, doi: 10.1007/BF00158429

    Nordlund, A. 1985, Solar Physics, 100, 209, doi: 10.1007/BF00158429

    work page doi:10.1007/bf00158429 1985

  23. [23]

    D., & Marshall, G

    Olivier, J., Johnson, W. D., & Marshall, G. D. 2008, Annals of Allergy, Asthma & Immunology, 100, 333, doi: 10.1016/S1081-1206(10)60595-9

    work page doi:10.1016/s1081-1206(10)60595-9 2008

  24. [24]

    Panda, J. P. 2021, arXiv e-prints, doi: 10.48550/arXiv.2111.07043

    Pith/arXiv arXiv doi:10.48550/arxiv.2111.07043 2021

  25. [25]

    2019, PyTorch: An Imperative Style, High-Performance Deep Learning Library, doi: 10.48550/arXiv.1912.01703

    Paszke, A., Gross, S., Massa, F., et al. 2019, PyTorch: An Imperative Style, High-Performance Deep Learning Library, doi: 10.48550/arXiv.1912.01703

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 2019

  26. [26]

    2011, Journal of Machine Learning Research, 12, 2825, doi: 10.5555/1953048.2078195

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825, doi: 10.5555/1953048.2078195

    doi:10.5555/1953048.2078195 2011

  27. [27]

    2020, pandas-dev/pandas: Pandas, doi: 10.5281/zenodo.3509134

    Reback, J., et al. 2020, pandas-dev/pandas: Pandas, doi: 10.5281/zenodo.3509134

    doi:10.5281/zenodo.3509134 2020

  28. [28]

    2018, The Astrophysical Journal, 859, 161, doi: 10.3847/1538-4357/aabba0

    Rempel, M. 2018, The Astrophysical Journal, 859, 161, doi: 10.3847/1538-4357/aabba0

    doi:10.3847/1538-4357/aabba0 2018

  29. [29]

    R., Warner, M., Keil, S

    Rimmele, T. R., Warner, M., Keil, S. L., et al. 2020, Solar Physics, 295, 172, doi: 10.1007/s11207-020-01736-7

    doi:10.1007/s11207-020-01736-7 2020

  30. [30]

    M., Kitiashvili, I

    Sadykov, V. M., Kitiashvili, I. N., Kosovichev, A. G., & Wray, A. A. 2021, The Astrophysical Journal, 909, 35, doi: 10.3847/1538-4357/abd9c7

    work page doi:10.3847/1538-4357/abd9c7 2021

  31. [31]

    2015, in ICLR, doi: 10.48550/arXiv.1409.1556

    Simonyan, K., & Zisserman, A. 2015, in ICLR, doi: 10.48550/arXiv.1409.1556

    Pith/arXiv arXiv doi:10.48550/arxiv.1409.1556 2015

  32. [32]

    1963, Monthly Weather Review, 91, 99, doi: 10.1175/1520-0493(1963)091⟨0099: GCEWTP⟩2.3.CO;2

    Smagorinsky, J. 1963, Monthly Weather Review, 91, 99, doi: 10.1175/1520-0493(1963)091⟨0099: GCEWTP⟩2.3.CO;2

    doi:10.1175/1520-0493(1963)091 1963

  33. [33]

    F., & Nordlund, ˚A

    Stein, R. F., & Nordlund, ˚A. 1998, The Astrophysical Journal, 499, 914, doi: 10.1086/305678 Szanda la, T. 2021, in Bio-inspired Neurocomputing, ed. A. K. Bhoi, P. K. Mallick, C.-M. Liu, & V. E. Balas (Singapore: Springer Singapore), 203–224, doi: 10.1007/978-981-15-5495-7 11

    doi:10.1086/305678 1998

  34. [34]

    2020, Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2 V¨ ogler, A., Shelyag, S., Sch¨ ussler, M., et al

    Virtanen, P., et al. 2020, Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2 V¨ ogler, A., Shelyag, S., Sch¨ ussler, M., et al. 2005, Astronomy & Astrophysics, 429, 335, doi: 10.1051/0004-6361:20041507

    doi:10.1038/s41592-019-0686-2 2020

  35. [35]

    Waidele, M., Zhao, J., & Kitiashvili, I. N. 2023, ApJ, 949, 99, doi: 10.3847/1538-4357/accae2

    doi:10.3847/1538-4357/accae2 2023

  36. [36]

    2020, Applied Sciences, 10, doi: 10.3390/app10051897

    Wang, Y., Li, Y., Song, Y., & Rong, X. 2020, Applied Sciences, 10, doi: 10.3390/app10051897

    work page doi:10.3390/app10051897 2020

  37. [37]

    N., & Kosovichev, A

    Mansour, N. N., & Kosovichev, A. G. 2015, arXiv e-prints, arXiv:1507.07999, doi: 10.48550/arXiv.1507.07999

    Pith/arXiv arXiv doi:10.48550/arxiv.1507.07999 2015

  38. [38]

    N., & Kosovichev, A

    Mansour, N. N., & Kosovichev, A. G. 2018, Realistic Simulations of Stellar Radiative MHD (Les Ulis: EDP Sciences), 39–62, doi: doi:10.1051/978-2-7598-2196-9.c004

    work page doi:10.1051/978-2-7598-2196-9.c004 2018

  39. [39]

    K., & Togashi, K

    Yamashita, R., Nishio, M., Do, R. K., & Togashi, K. 2018, Insights into Imaging, 9, 611, doi: 10.1007/s13244-018-0639-9 45

    doi:10.1007/s13244-018-0639-9 2018

  40. [40]

    2014, in International Conference on Learning Representations, doi: 10.48550/arXiv.1412.6856

    Oliva, A. 2014, in International Conference on Learning Representations, doi: 10.48550/arXiv.1412.6856

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6856 2014