Pith. sign in

REVIEW 3 major objections 4 minor 18 references

Cluster-Weighted Training of Deep Surrogate Models for Subgrid Turbulent Transport

T0 review · 3 major / 4 minor · reviewed 2026-07-10 · grok-4.5

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

Pith's one-line read Cluster-weighted training cuts CNN error on solar Reynolds stresses by 34 percent and raises R^{2} from 0.54 to 0.80.

desk verdict Solid empirical lift from K-Means error reweighting on a solar Reynolds-stress CNN; the 34% MSE / R^{2} gains are real for this setup, but rest on an unablated (T0, α) schedule and still lack closed-loop physics checks. read the letter →

arxiv 2607.07925 v1 pith:CMHDCEWS submitted 2026-07-08 astro-ph.SR physics.flu-dyn

Rimsha Hameed Syeda , Dustin Kempton , Viacheslav Sadykov , Irina Kitiashvili , Rafal Angryk This is my paper · ORCID
classification astro-ph.SRphysics.flu-dyn
keywords surrogatemodelingsubgridturbulenceReynoldsstresstensorcluster-weightedtrainingK-Means3-DCNNsolarconvectionquietSun
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

Solar-atmosphere simulations cannot resolve every turbulent eddy, so they need a fast surrogate that estimates the missing Reynolds-stress tensor from large-scale density and velocity. Earlier work already showed that a 3-D convolutional network beats classical gradient and Smagorinsky closures, yet the network still errs systematically in certain flow regimes. This paper partitions the input space with K-Means, measures the residual error of each cluster after a short uniform warm-up, and reweights the loss so that high-error clusters receive more gradient attention. The resulting model recovers both the peaks and the heavy tails of the stress distributions more faithfully, lowers average mean-squared error by more than a third, and lifts the coefficient of determination from 0.54 to 0.80. The same idea is offered as a general remedy for regression problems whose error surface is spatially heterogeneous.

What carries the argument

Cluster-weighted loss: after ten epochs of uniform training, samples belonging to each K-Means cluster are multiplied by a scalar proportional to that cluster’s residual MSE; the reweighted network is then fine-tuned to completion.

What would settle it

Retrain the identical CNN architecture on the same solar cubes using a non-cluster importance scheme (for example, uncertainty-based or residual-based sample weights) and check whether the 34 % MSE reduction and the R^{2} jump to 0.80 still appear; if they do not, the clustering step is essential.

Watch

Extended reading notes

Core claim

When a 3-D CNN that predicts the six Reynolds-stress components is retrained with a loss reweighted by the mean squared error of five K-Means clusters (weights wk = 1 + 3 · MSEk / max MSEj), average MSE falls from 4.89 × 10^18 to 3.01 × 10^18 and average R^{2} rises from 0.54 to 0.80 relative to the identical architecture trained with ordinary MSE.

Load-bearing premise

That the cluster-wise errors measured after only ten epochs of ordinary training give a stable, non-overfitting signal for how much each sample should be weighted thereafter.

Editorial extensions

If this is right

  • Lower-resolution StellarBox runs can replace classical subgrid closures with the cluster-weighted CNN and still recover high-resolution Reynolds-stress statistics.
  • Off-diagonal stress components, previously only modestly improved by deep networks, now show large error reductions once high-error clusters are emphasized.
  • The same reweighting recipe can be applied to any regression task whose input space contains regions of systematically higher residual.
  • Inference cost remains identical to the baseline CNN because only the training loss, not the architecture, is changed.

Reading between the lines

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

  • Because the largest gains concentrate between –1 Mm and the photosphere, the method may be especially useful for near-surface magnetoconvection where gradients are sharpest.
  • The modest 15–20 % training overhead suggests the technique could be inserted into other large-eddy-simulation pipelines without redesigning the network.
  • If magnetic-field channels were added to the input features, the same cluster-weighting loop might further reduce errors in magnetized turbulence.
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 / 4 minor

Summary. The paper introduces cluster-weighted training for a 3D CNN surrogate that predicts the six Reynolds-stress components τ ij from local macroscopic velocities and density, using high-resolution StellarBox quiet-Sun cubes. After a short uniform-MSE warm-up (T0=10 epochs), K-Means (K=5) partitions of the 82-dimensional input features supply permanent sample weights wk=1+α·(MSEk/max MSEj) with α=3; the CNN is then retrained under the weighted loss. Relative to the identical architecture trained with uniform MSE, average MSE falls from 4.89 imes10^18 to 3.01 imes10^18 (≈34 %) and average R² rises from 0.54 to 0.80 (Table II), with clearer PDF tails, tighter error histograms and scatter plots (Figs. 5–7) and localized gains in selected clusters and height ranges (Fig. 8). Limited HAC and physics-baseline comparisons are also shown.

Significance. If the reported gains prove robust, the work supplies a practical, architecture-preserving recipe for improving deep surrogates of anisotropic subgrid transport in solar and stellar convection-zone models. The multi-metric evaluation (RMSE per component, PDFs, height-resolved relative improvement) and the explicit algorithmic description (Fig. 4) are strengths that aid reproducibility. The technique is potentially transferable to other regression problems whose error landscape is spatially heterogeneous, which is a useful methodological contribution beyond the immediate solar-physics application.

major comments (3)
  1. [Section V-D / Fig. 4] Section V-D and the algorithm of Fig. 4 freeze the cluster weights after only T0=10 epochs of uniform training and employ a single hand-chosen scaling α=3. No ablation of T0 or α, nor a control that reweights by a random or magnitude-matched schedule of the same strength, is reported. Because Table II attributes the 34 % MSE / R² gains specifically to cluster-weighted training, these untested choices are load-bearing; without them it remains possible that any aggressive reweighting of the high-error tail after a short warm-up would produce comparable lifts.
  2. [Tables I–II] Tables I–II report only point estimates of RMSE, average MSE and R². Despite the authors having already generated 20 K-Means seeds and five representative clusterings, no standard deviations, bootstrap intervals or significance tests accompany the headline numbers. The quantitative claim of a 34 % reduction and an R² jump from 0.54 to 0.80 therefore lacks a statistical foundation.
  3. [Section VI-A / Table I] Section VI-A and Table I compare HAC-based weighting only on a 2.5 % random subset of the training data (memory-limited) and with a different power-law scheme. The resulting under-performance relative to full-data K-Means cannot be interpreted as a fair head-to-head of clustering methods, yet the text presents it as supporting evidence for the superiority of K-Means.
minor comments (4)
  1. [Abstract] Abstract and conclusion state “34 % lower MSE” without citing the precise average-MSE values of Table II; a parenthetical reference would improve traceability.
  2. [Fig. 2] Figure 2a reports cluster-wise MSE in cm⁴ s⁻⁴ while the later tables use cm² s⁻²; units should be harmonized or explicitly converted.
  3. [Section III] The signed-log transform applied to off-diagonal targets is mentioned only briefly; a short formula or reference would help readers reproduce the preprocessing pipeline.
  4. [Figs. 5–8] Several figure panels (especially Fig. 7 scatter plots) lack axis labels or color-bar units; adding them would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: cluster-weighted improvement is an empirical reweighting result measured on held-out data, not forced by definition or self-citation chain.

full rationale

The paper's central claim (34% MSE reduction and R^{2} lift from 0.54 to 0.80) is produced by an explicit experimental procedure: unsupervised K-Means on input features only, a short uniform-MSE warm-up (T0=10), freezing of per-cluster mean errors into fixed sample weights wk=1+α·(MSEk/max MSEj) with α=3, then full retraining under the weighted loss, followed by evaluation on a held-out test set (Tables I-II, Figs. 5-8, Algorithm in Fig. 4 / Sec. V-D). None of these steps reduces the reported metrics to an identity or to a fitted parameter by construction; the weights are a heuristic derived from early errors and the final numbers are independent measurements. Self-citations to the authors' prior work [3] supply only the CNN architecture, data pipeline and uniform-loss baseline; they are not load-bearing for the weighting gain itself, which is newly measured here against that baseline and against physics-based models. No uniqueness theorem, ansatz smuggling, or renaming of a known result appears. The procedure is therefore self-contained empirical ML methodology, not circular derivation.

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

The central claim rests on a handful of free hyper-parameters chosen by silhouette or hand, standard definitions of Reynolds stress and MSE, and the modeling assumption that local 3×3×3 velocity-density patches plus unsupervised clustering of those patches are sufficient to capture and correct error heterogeneity. No new physical entities are postulated.

free parameters (4)
  • number of clusters K = 5
    Fixed at 5 after maximizing average silhouette width on the 82-dimensional feature space; directly controls the granularity of the reweighting.
  • weight scaling factor α = 3
    Hand-chosen multiplier that amplifies high-error clusters in the loss; set to 3 without reported sensitivity study.
  • initial uniform-training epochs T0 = 10
    Number of epochs used to obtain the baseline errors that define the cluster weights; fixed at 10.
  • Adam learning rate and batch size = 0.001 / 128
    Standard optimizer settings (0.001, 128) inherited from prior work and not re-tuned for the weighted regime.
assumptions (4)
  • domain assumption Reynolds stress is defined by τij = <ui uj> − <ui><uj> averaged over the 4×4×4 sub-cube.
    Standard continuum definition used both to generate labels and as the prediction target (Eq. 1).
  • domain assumption A 3×3×3 neighborhood of velocity components plus central density is a sufficient feature set for predicting the six stress components.
    Inherited from the authors’ prior pipeline [3] and never re-justified here.
  • ad hoc to paper K-Means partitions of the input feature space align with regions of systematically higher prediction error, so that reweighting those partitions improves generalization.
    Core methodological premise of Section V; supported only by the post-hoc error analysis in Fig. 2.
  • domain assumption Non-overlapping spatial sampling every third sub-cube and 5-minute temporal cadence yield statistically independent training samples.
    Stated in Section III to justify random train/validation/test splits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cluster-Weighted Training of Deep Surrogate Models for Subgrid Turbulent Transport." pith.science (2026). https://pith.science/paper/CMHDCEWS

@misc{pith2026260707925,
  author       = {Pith},
  title        = {Pith review of: Cluster-Weighted Training of Deep Surrogate Models for Subgrid Turbulent Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMHDCEWS}},
  note         = {Machine review of arXiv:2607.07925}
}
read the original abstract

Turbulence in the solar interior and atmosphere plays a crucial role in energy transport, yet modeling its subgrid-scale effects remains a major challenge. This study leverages machine learning (ML) models to predict components of the Reynolds stress tensor using high-resolution StellarBox simulations of the quiet Sun. Previously, we have compared a Multi-Layer Perceptron (MLP) and a 3D Convolutional Neural Network (CNN) against physics-based baselines to achieve a lower Mean Squared Error (MSE) and better generalization across various heights and depths in the solar atmosphere. To enhance learning, in this work, we investigate cluster-weighted training using K-Means and Hierarchical Agglomerative Clustering (HAC). By weighing the loss function based on cluster-specific prediction errors, we direct the model's attention to high-error regions. It significantly improves CNN performance, achieving 34% lower MSE and a significantly higher R2 score indicating that integrating deterministic clustering with ML is a promising technique for modeling subgrid turbulence, in particular, and regression in diverse environments, in general.

Figures

Figures reproduced from arXiv: 2607.07925 by the authors.

Figure 1
Figure 1. Illustration of vertical velocities in the simulations ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Cluster-wise MSE (cm4/ s 4 ) for each τij component using baseline CNN predictions. (b) Distribution of clusters as a function of height. to investigate the role of clustering quality and initialization strategies: • Standard KMeans: selects k initial centroids uniformly at random from the data; • KMeans++: selects centroid using probability propor￾tional to its squared distance from the closest existing centroi… view at source ↗
Figure 3
Figure 3. Distribution of inertia across random states using KMeans++ and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Cluster-Weighted Training method is well-suited for uncovering hierarchical struc￾ture in the data, but is computationally expensive. Due to memory constraints, we were only able to perform cluster-weighted training using HAC on a subset compris￾ing 2.5% of the full tr…
Figure 5
Figure 5. Figure 5: Distributions of components τvv (a) and τuw (b). Blue curves show the original target data, orange shows the predictions of the CNN, and gray shows the predictions of the cluster-weighted CNN (ClCNN). Model Average MSE Average R2 CNN (uniform loss) 4.89 × 1018 0.54 CNN…
Figure 6
Figure 6. Figure 6: Histogram of absolute prediction errors for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of the predictions by the cluster-weighted CNN and the baseline CNN with the actual values of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Cluster-wise median relative improvement of MSE (a) and relative improvements as a function of height (b) across all [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

  1. [1]

    Reynolds Stress Modeling Using Data Driven Machine Learning Algorithms

    J. P. Panda, “Reynolds Stress Modeling Using Data Driven Machine Learning Algorithms,”arXiv e-prints, 2021. [Online]. Available: https://doi.org/10.48550/arXiv.2111.07043

    work page Pith review arXiv doi:10.48550/arxiv.2111.07043 2021

  2. [2]

    Physics-informed machine learning for modeling turbulence in supernovae,

    P. I. Karpov, C. Huang, I. Sitdikov, C. L. Fryer, S. Woosley, and G. Pilania, “Physics-informed machine learning for modeling turbulence in supernovae,”The Astrophysical Journal, vol. 940, no. 1, p. 26, 2022. [Online]. Available: https://doi.org/10.3847/1538-4357/ac88cc

    work page doi:10.3847/1538-4357/ac88cc 2022

  3. [3]

    Developing machine learning models of subgrid turbulent transport for quiet sun 3d radiative hydrodynamic simulations,

    R. H. Syeda, D. Kempton, V . Sadykov, I. Kitiashvili, and R. Angryk, “Developing machine learning models of subgrid turbulent transport for quiet sun 3d radiative hydrodynamic simulations,” SSRN Preprint No. 5236395, 2025. [Online]. Available: https://ssrn.com/abstract=5236395

    work page 2025

  4. [4]

    Using Realistic MHD Simu- lations for Modeling and Interpretation of Quiet-Sun Observations with the Solar Dynamics Observatory Helioseismic and Magnetic Imager,

    I. N. Kitiashvili, S. Couvidat, and A. Lagg, “Using Realistic MHD Simu- lations for Modeling and Interpretation of Quiet-Sun Observations with the Solar Dynamics Observatory Helioseismic and Magnetic Imager,” The Astrophysical Journal, vol. 808, no. 1, p. 59, Jul. 2015. 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 log10|Predicted| - log10|Target| 0.00 0.25 0.50 0....

    work page 2015

  5. [5]

    Simulations of Stellar Magnetoconvection using the Radiative MHD Code `StellarBox'

    A. A. Wray, K. Bensassi, I. N. Kitiashvili, N. N. Mansour, and A. G. Kosovichev, “Simulations of Stellar Magnetoconvection using the Radiative MHD Code ‘StellarBox’,”arXiv e-prints, p. arXiv:1507.07999, Jul. 2015. [Online]. Available: https://doi.org/10. 48550/arXiv.1507.07999

    work page Pith review arXiv 2015

  6. [6]

    Realistic Simulations of Stellar Radiative MHD,

    A. A. Wray, K. Bensassiy, I. N. Kitiashvili, N. N. Mansour, and A. G. Kosovichev, “Realistic Simulations of Stellar Radiative MHD,” inVariability of the Sun and Sun-Like Stars: from Asteroseismology to Space Weather, J. P. Rozelot and E. S. Babayev, Eds., 2018, p. 39. [Online]. Available: https://doi.org/10.1007/978-3-319-77199-1 4

    work page doi:10.1007/978-3-319-77199-1 2018

  7. [7]

    Solar Convection,

    A. Nordlund, “Solar Convection,”Solar Physics, vol. 100, p. 209, Oct. 1985. (a) (b) Fig. 8. Cluster-wise median relative improvement of MSE (a) and relative improvements as a function of height (b) across allτ

    work page 1985

  8. [8]

    Revisiting turbulent properties of solar convection with 3D radiative hydrodynamic modeling

    I. N. Kitiashvili and A. A. Wray, “Revisiting turbulent properties of solar convection with 3D radiative hydrodynamic modeling,”arXiv e-prints, p. arXiv:2502.00974, Feb. 2025

    work page Pith review arXiv 2025

Show all 18 references
  1. [9]

    General Circulation Experiments with the Primitive Equations,

    J. Smagorinsky, “General Circulation Experiments with the Primitive Equations,”Monthly Weather Review, vol. 91, no. 3, p. 99, Jan. 1963. [Online]. Available: https://doi.org/10.1175/1520-0493(1963)091⟨0099: GCEWTP⟩2.3.CO;2

    1963 doi

  2. [10]

    A dynamic subgrid-scale eddy viscosity model,

    M. Germano, U. Piomelli, P. Moin, and W. H. Cabot, “A dynamic subgrid-scale eddy viscosity model,”Physics of Fluids A, vol. 3, no. 7, pp. 1760–1765, Jul. 1991. [Online]. Available: https://doi.org/10.1063/1.857955

    1991 doi

  3. [11]

    Reynolds averaged turbulence modelling using deep neural networks with embedded invariance,

    J. Ling, A. Kurzawski, and J. Templeton, “Reynolds averaged turbulence modelling using deep neural networks with embedded invariance,” Journal of Fluid Mechanics, vol. 807, pp. 155–166, 2016. [Online]. Available: https://doi.org/10.1017/jfm.2016.615

    2016 doi

  4. [12]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational Physics, vol. 378, pp. 686–707, 2019

    2019

  5. [13]

    Nonzero Phase Shifts of Acoustic Waves in the Lower Solar Atmosphere Measured from Realistic Simulations and Their Role in Local Helioseismology,

    M. Waidele, J. Zhao, and I. N. Kitiashvili, “Nonzero Phase Shifts of Acoustic Waves in the Lower Solar Atmosphere Measured from Realistic Simulations and Their Role in Local Helioseismology,”The Astrophysical Journal, vol. 949, no. 2, p. 99, Jun. 2023

    2023

  6. [14]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,”Journal of Machine Learn...

    2011 doi

  7. [15]

    Data clustering: 50 years beyond k-means,

    A. K. Jain, “Data clustering: 50 years beyond k-means,”Pattern Recog- nition Letters, vol. 31, no. 8, pp. 651–666, 2010

    2010

  8. [16]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: a graphical aid to the interpretation and validation of cluster analysis,”Journal of Computational and Applied Mathematics, vol. 20, pp. 53–65, 1987

    1987

  9. [17]

    k-means++: The advantages of careful seeding,

    D. Arthur and S. Vassilvitskii, “k-means++: The advantages of careful seeding,” inProceedings of the eighteenth annual ACM-SIAM sym- posium on Discrete algorithms. Society for Industrial and Applied Mathematics, 2007, pp. 1027–1035

    2007

  10. [18]

    Small-scale Dynamo Simulations: Magnetic Field Am- plification in Exploding Granules and the Role of Deep and Shallow Recirculation,

    M. Rempel, “Small-scale Dynamo Simulations: Magnetic Field Am- plification in Exploding Granules and the Role of Deep and Shallow Recirculation,”The Astrophysical Journal, vol. 859, no. 2, p. 161, Jun. 2018

    2018

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.