Pith. sign in

REVIEW 2 major objections 6 minor 59 references

Equivariant nonlocal networks beat data-augmented ones for subgrid stress at half the parameters, while pointwise models barely beat Clark.

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

2026-07-30 19:13 UTC pith:OC2A76DV

load-bearing objection Clean a-priori bake-off: equivariant nonlocal SGS nets beat augmented CNNs at half the parameters; pointwise models do not beat Clark—useful design guidance, not yet a solver result. the 2 major comments →

arxiv 2607.26850 v1 pith:OC2A76DV submitted 2026-07-29 physics.flu-dyn physics.comp-ph

Rotational equivariance and locality in data-driven subgrid-scale closures

Ryley McConkey , Julia Balla , Elyssa Hofgard , Tess Smidt , Abigail Bodner This is my paper
classification physics.flu-dyn physics.comp-ph
keywords large eddy simulationsubgrid-scale modellingdata-driven turbulence closureequivariant neural networksrotational octahedral groupnonlocalitychannel flowparameter efficiency
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.

Large-eddy simulation needs a cheap model for the stress that unresolved scales exert on the resolved flow. This paper asks whether building the discrete rotational symmetry of a Cartesian grid into the network helps that model, and whether the network must look at a neighbourhood rather than a single point. On filtered channel-flow DNS at a realistic filter width, four architectures are compared at matched parameter counts: pointwise and nonlocal, each with and without architectural equivariance. Non-equivariant nets pick up a little rotational symmetry from the turbulence data alone, more so in near-isotropic regions. The equivariant nonlocal model wins every generalization test—held-out space and time, swapped anisotropy, higher Reynolds number—at roughly half the parameters of its augmented non-equivariant twin, and it is more data-efficient. Pointwise nets, equivariant or not, do not beat the classical Clark formula. The practical claim is that both equivariance and nonlocality matter for data-driven closures under realistic data, parameter, and filter budgets.

Core claim

At matched parameter counts on turbulent channel flow with a realistic filter ratio, an equivariant nonlocal architecture attains the highest a-priori correlation on every generalization axis (spatiotemporal, anisotropy, Reynolds number) at about half the parameter count of a data-augmented non-equivariant CNN, while both pointwise architectures fail to improve on the analytical Clark baseline; the benefit of equivariance grows with receptive field, and the equivariant model is also more data-efficient.

What carries the argument

Matched-parameter comparison of four closures that map the filtered velocity-gradient tensor to the deviatoric subgrid stress: an SO(3)-equivariant strain-rate eigenframe MLP, a plain MLP with octahedral augmentation, a steerable group-equivariant CNN over the rotational octahedral group O, and a plain 3-D CNN with the same augmentation—plus the non-trainable Clark gradient model as baseline.

Load-bearing premise

That ranking models by how well they match filtered DNS stress a priori is enough to decide which architecture is better once the model is dropped into a live large-eddy simulation, where stability and dissipation matter.

What would settle it

Insert the trained ESCNN and the matched-parameter augmented CNN into the same LES solver on channel flow (and a second flow) and check whether the a-priori correlation ordering survives in a-posteriori statistics such as mean profiles, spectra, and long-time stability.

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

If this is right

  • Deployable data-driven SGS closures at realistic filter widths should be nonlocal; pointwise maps from the local velocity gradient are information-limited and do not beat Clark.
  • Architectural equivariance under the discrete octahedral group is worth the implementation cost in the low-parameter, low-data regime typical of CFD closures.
  • The value of equivariance increases with receptive field, so larger-stencil or multi-scale equivariant closures should widen the gap further.
  • Training on more isotropic turbulence alone already imparts partial equivariance, so augmentation budgets can be reduced when the training region is near-isotropic.

Where Pith is reading between the lines

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

  • If a-posteriori tests preserve the ranking, production LES codes could adopt steerable octahedral convolutions as a default inductive bias for learned SGS tensors rather than relying on heavy data augmentation.
  • The same matched-parameter protocol could decide whether reflection equivariance (full Oh) or continuous SO(3) steerable layers add further gains once the discrete rotational symmetry is already enforced.
  • Because the gap appears only when the model has spatial extent, hybrid schemes that keep a cheap local base model and learn only a nonlocal residual may capture most of the benefit at still lower cost.

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

2 major / 6 minor

Summary. The manuscript studies whether architectural rotational equivariance under the discrete octahedral group O improves data-driven SGS closures relative to data-augmented non-equivariant models, and how that interacts with locality. On JHTDB channel flow at Re_τ=1000 and 5200 (box-filtered at ratio 4, Δ+≈40), the authors compare pointwise (eigenframe MLP vs augmented MLP) and nonlocal (steerable ESCNN vs augmented 3D CNN) models at matched parameter counts, with a Clark analytical baseline. Inputs/outputs are nondimensionalized by local G and Δ²G². They report (i) modest implicit equivariance learned from non-augmented turbulence data, stronger in more isotropic regions; (ii) ESCNN highest ρ_τ on spatiotemporal, anisotropy, and Reynolds generalization at roughly half the CNN parameter count; (iii) pointwise models not beating Clark; (iv) equivariance benefit growing with receptive field. Evaluation is a priori only.

Significance. The work addresses a timely, practical question in data-driven LES: whether the architectural cost of discrete rotational equivariance is repaid at realistic filter ratios, parameter budgets, and dataset sizes. Strengths include a controlled matched-parameter bake-off, three generalization axes, careful Galilean-invariant nondimensionalization that enables Re transfer, an analytical Clark baseline, equivariance-error and commutative-diagram diagnostics, public JHTDB extraction, and stated code availability. The finding that nonlocality is necessary at Δ+≈40 and that equivariance helps more once the model is nonlocal is useful guidance for the community and aligns with broader scientific ML patterns without overselling continuous SO(3). Within a priori scope the comparison is among the cleaner ones available for tensorial SGS mappings.

major comments (2)
  1. [§4 Conclusion; Abstract; Table 3] §4 and the Abstract/Conclusion recommendation: the central practical claim that an equivariant nonlocal architecture is advantageous for SGS modelling rests entirely on a priori ρ_τ (Table 3, Figs. 5–7). The manuscript correctly flags that closures were not run inside an LES solver and that stability, dissipation, and discretization interaction are untested. Because the conclusion still frames ESCNN as the recommended deployable choice, the recommendation should be explicitly scoped to a priori stress matching, or a minimal a posteriori check (e.g., frozen-coefficient channel LES energy spectra/dissipation or a short solver-in-the-loop stability diagnostic) should be added. Without that, the transfer from Table 3 rankings to “practical data-driven SGS modelling” remains an assumption, not a result.
  2. [§2.5 Training procedure; Table 3; §3.2–3.3] §2.5 / §3: all configurations use a single random seed, with trends across size and data sweeps offered in lieu of uncertainty. For the load-bearing claim that ESCNN beats the augmented CNN on every generalization cell at half the parameters (Table 3: e.g. near-wall spatiotemporal 0.760 vs 0.700; channel-center 0.874 vs 0.814), at least the primary matched-parameter pairs should be repeated over a few seeds or report validation-loss variability. Single-seed point estimates are otherwise hard to distinguish from training noise at the reported correlation gaps of ~0.05–0.06.
minor comments (6)
  1. [§3.2; Figures 5–6] Figs. 5–6 wall-clock panels: the text acknowledges that steerable convolutions are less-optimized research code versus production CNN kernels, and that eigenframe cost is dominated by eigendecomposition. Consider moving wall-clock from a primary efficiency axis to a clearly caveated secondary metric, or reporting FLOPs/parameter throughput, so parameter- and data-efficiency (the cleaner comparisons) remain the headline.
  2. [§2.2] §2.2: the choice to enforce O but not O_h (reflections) is stated and deferred; a one-sentence note on whether channel-flow statistics or the filtered equations make reflection equivariance likely to matter for SGS would help readers judge urgency of that extension.
  3. [Table 1; §3.3] Table 1 / §2.4.3: barycentric coordinates usefully document anisotropy; adding the corresponding region-averaged |τ^d| or G statistics would help interpret why channel-center Re generalization correlations exceed spatiotemporal ones (text already notes higher C_3c at Re_τ=5200).
  4. [Abstract; throughout] Typos/spacing: several compounded words appear without spaces in the compiled text (e.g. “abouttheroleofrotationalequivariance”, “evaluatedatmatchedparametercounts”). Sweep the PDF for missing spaces after copy-editing or LaTeX line-break artifacts.
  5. [§2.3.3; Appendix A] Appendix A depth sweep supports fixing nonlocal depth at 4; a brief cross-reference in §2.3.3 would make the main-text depth choice easier to find.
  6. [§1 Introduction] Related work: Agdestein & Sanderse (2025, 2026) are cited appropriately for discrete vs continuous symmetry; ensure the 2026 arXiv comparison paper is distinguished from the present contribution so novelty on matched equivariant/nonlocal SGS bake-offs is clear.

Circularity Check

0 steps flagged

No significant circularity: empirical bake-off against external JHTDB targets and a fixed Clark baseline, not self-defined predictions.

full rationale

This paper is a controlled a-priori architecture comparison (equivariant vs data-augmented non-equivariant; pointwise vs nonlocal) trained and scored on explicitly box-filtered Johns Hopkins channel-flow DNS, with held-out spatiotemporal, anisotropy, and Reynolds-number splits and a parameter-free analytical Clark baseline. Correlation ρ_τ and equivariance error are standard external diagnostics: equivariant nets have zero equivariance error by architectural construction (a design property, not a claimed empirical discovery), while non-equivariant nets are measured against the same commutative residual on real fields. Self-citations to the authors’ prior super-resolution / implicit-augmentation work supply related background observation only; they do not define the SGS targets, force the Table 3 rankings, or substitute for the matched-parameter generalization results. No step reduces a claimed prediction to a fitted input by construction, imports a uniqueness theorem from the same authors, or renames a known closed-form result as a derived finding. The acknowledged a-priori/a-posteriori gap is a scope limitation, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

Central claims are empirical rankings under stated modelling choices, not theorems. Load-bearing assumptions are standard LES/SGS modelling choices (gradient-only Galilean-invariant input, deviatoric stress target, box filter ratio 4, discrete group O not Oh) plus ML training protocol (single seed, Adam, MSE on normalized tensors). No new physical entities. Free parameters are ordinary network hyperparameters and the hand-chosen filter/geometry settings that define the task.

free parameters (4)
  • Network width/depth per architecture (MLP depth 1–8 width 64; CNN channels; ESCNN regular-rep copies; fixed nonlocal dep = Depth 4 nonlocal; widths spanning ~10^3–10^6 params
    Swept and selected on validation loss; depth 4 chosen because it won at matched parameter count (Appendix A). Rankings depend on these operating points.
  • Learning rate and batch size (per architecture, fixed across size sweep)
    Chosen so largest model trains stably (§2.5); not cross-validated publicly in detail; can shift which model wins at the margin.
  • Filter ratio / Δ+ and box extraction geometry = ratio 4, Δ+≈40, L+=640
    Downsampling ratio 4, Δ+≈40, y+=321 near-wall centering, 64^3→16^3→14^3 interior define the SGS task difficulty and locality; conclusions are for this realistic but single filter setting.
  • Input/output normalization scale G and Δ^2 G^2 = Ĝ=G_ij/G, τ̂^d=τ^d/(Δ^2 G^2)
    Hand-specified nondimensionalization (Eqs. 3–4) following mixing-length scaling; enables Re comparison but shapes the learned map and reported losses.
axioms (5)
  • domain assumption Discretized filtered NS on a uniform Cartesian grid are equivariant under the rotational octahedral group O (24 elements), not full continuous SO(3)/O(3).
    §2.2; justifies enforcing O-equivariance and octahedral augmentation rather than continuous rotations.
  • domain assumption Deviatoric SGS stress is a function of the filtered velocity gradient (and its local stencil for nonlocal models); filter width enters only via normalization, not as an extra input.
    §2.1; standard Galilean-invariant constitutive assumption following Park & Choi / Prakash et al.
  • domain assumption A priori MSE/correlation on explicitly filtered DNS is a meaningful ranking metric for SGS model quality.
    §2.6, §3; standard in the literature but known to be incomplete vs a posteriori LES.
  • ad hoc to paper Octahedral data augmentation is the fair non-architectural baseline for teaching O-equivariance to ordinary MLPs/CNNs.
    §2.5; comparison design choice—other regularizers or approximate equivariance methods are not tested.
  • ad hoc to paper Single-seed training trends across size/data sweeps suffice without seed-averaged error bars.
    §2.5 explicitly trades multi-seed variance for breadth of sweeps.

pith-pipeline@v1.2.0-daily-grok45 · 28774 in / 3748 out tokens · 73626 ms · 2026-07-30T19:13:54.916370+00:00 · methodology

0 comments
read the original abstract

Data-driven subgrid-scale closures for large eddy simulation are of significant interest in many engineering and geoscience applications. In this context, several important questions remain about the role of rotational equivariance as an inductive bias for learned tensorial mappings. We investigate whether equivariance improves accuracy, parameter efficiency, and generalization for subgrid-scale modelling at realistic filter ratios. For turbulent channel flow, we compare data-augmented non-equivariant architectures to those with equivariance as an inductive bias. We compare both pointwise and nonlocal versions of these two model classes. All models are evaluated at matched parameter counts across spatiotemporal, anisotropy, and Reynolds number generalization. We show that non-augmented models learn a small degree of equivariance directly from turbulence data, especially when that data is more isotropic. The equivariant nonlocal architecture attains the highest correlation coefficient on every generalization test at approximately half the parameter count of its non-equivariant counterpart, while the pointwise architectures do not improve on the analytical Clark baseline. Additionally, the equivariant model is more data-efficient than a non-equivariant model. The benefit of equivariance grows with the receptive field of the model, indicating that equivariance and nonlocality are both useful for the subgrid-scale closure task at realistic dataset size, parameter counts, and filter size.

Figures

Figures reproduced from arXiv: 2607.26850 by Abigail Bodner, Elyssa Hofgard, Julia Balla, Ryley McConkey, Tess Smidt.

Figure 1
Figure 1. Figure 1: Extraction box locations in the two channels. Left: cross-section of the 𝑅𝑒𝜏 = 1000 channel. Right: cross-section of the 𝑅𝑒𝜏 = 5200 channel. Near-wall and channel-center boxes are drawn to scale on each channel. Center: viscous-scaled box side length, 𝐿+ = 640, common to all boxes across both Reynolds numbers. The near-wall boxes are centered at a matched wall distance 𝑦 + = 321 at both Reynolds numbers; t… view at source ↗
Figure 2
Figure 2. Figure 2: (a) spatial and (b) temporal split of the extraction boxes into training (𝑧0 , 𝑧1 , 𝑧2 ), validation (𝑧3 ), and test (𝑧4 ) columns. The split is applied jointly in space and time, with no overlap between the three sets. test sets are disjoint from the training set in both space and time [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Equivariance error of the non-equivariant pointwise MLP and CNN, with and without octahedral data augmentation, as a function of the number of training points. Results are shown for both the near-wall (blue) and channel-center (red) regions of the 𝑅𝑒𝜏 = 1000 channel. Equivariance error is evaluated on the spatiotemporal validation set for each training region [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Commutative diagram for the non-equivariant CNNs trained without data augmentation, trained/evaluated on the near-wall (a) and channel-center (b) regions of the 𝑅𝑒𝜏 = 1000 channel. The input panels show the dimensionless 𝐺̂ 33 component of the velocity gradient, and the output panels show the dimensionless ̂𝜏 d 33 component of the deviatoric SGS stress, on an 𝑥𝑦 slice at 𝑧 = 7 of the 143 interior. The diag… view at source ↗
Figure 5
Figure 5. Figure 5: Validation loss for the pointwise models trained on the 𝑅𝑒𝜏 = 1000 near-wall dataset, as a function of (a) the number of trainable parameters, (b) the inference wall-clock time per 10M prediction points, and (c) the training data fraction. (a) and (b) show the equivariant eigenframe model and the non-equivariant pointwise multilayer perceptron trained with octahedral data augmentation. (c) additionally sho… view at source ↗
Figure 6
Figure 6. Figure 6: Validation loss for the nonlocal models trained on the 𝑅𝑒𝜏 = 1000 near-wall dataset, as a function of (a) the number of trainable parameters, (b) the inference wall-clock time per 10M prediction points, and (c) the training data fraction. Panels (a) and (b) show the equivariant ESCNN and the non-equivariant CNN trained with octahedral data augmentation. Panel (c) additionally shows the non-equivariant CNN … view at source ↗
Figure 7
Figure 7. Figure 7: Magnitude of the predicted deviatoric SGS stress on an 𝑥𝑦 slice at 𝑧 = 7 of the 143 interior, for the Clark model and the four trained models, compared against the filtered DNS, on the three generalization tests of Section 2.4.4. Rows correspond to the spatiotemporal, anisotropy, and Reynolds number generalization tests. Columns correspond to the filtered DNS target, the Clark model, the Eigenframe MLP, th… view at source ↗
Figure 8
Figure 8. Figure 8: Validation loss on the 𝑅𝑒𝜏 = 1000 near-wall validation set for the non-equivariant, data-augmented CNN as a function of (a) channel width and (b) total number of parameters, with one curve per depth ∈ {2, 3, 4, 8}. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work the author(s) used generative AI in order to assist with proof… view at source ↗
Figure 9
Figure 9. Figure 9: Validation loss on the 𝑅𝑒𝜏 = 1000 near-wall validation set for the equivariant ESCNN as a function of (a) number of regular representation copies per layer and (b) total number of parameters, with one curve per depth ∈ {2, 3, 4, 8}. Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J.P., Kornbluth, M., Molinari, N., Smidt, T.E., Kozinsky, B., 2022. E(3)-equivariant graph neural networks for data-eff… 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

59 extracted references · 6 canonical work pages

  1. [1]

    GENERAL CIRCULATION EXPERIMENTS WITH THE PRIMITIVE EQUATIONS: I

    Smagorinsky, Joseph. GENERAL CIRCULATION EXPERIMENTS WITH THE PRIMITIVE EQUATIONS: I. THE BASIC EXPERIMENT. Monthly Weather Review. 1963. doi:10.1175/1520-0493(1963)091<0099:GCEWTP>2.3.CO;2

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

  2. [2]

    Journal of Fluid Mechanics , author=

    Evaluation of subgrid-scale models using an accurately simulated turbulent flow , volume=. Journal of Fluid Mechanics , author=. 1979 , pages=. doi:10.1017/S002211207900001X , number=

    work page doi:10.1017/s002211207900001x 1979

  3. [3]

    Theoretical and Computational Fluid Dynamics , year =

    Vreman, Bert and Geurts, Bernard and Kuerten, Hans , title =. Theoretical and Computational Fluid Dynamics , year =. doi:10.1007/BF00639698 , url =

    work page doi:10.1007/bf00639698

  4. [4]

    Turbulence Modeling in the Age of Data

    Duraisamy, Karthik and Iaccarino, Gianluca and Xiao, Heng. Turbulence Modeling in the Age of Data. Annual Review of Fluid Mechanics. 2019. doi:https://doi.org/10.1146/annurev-fluid-010518-040547

    doi:10.1146/annurev-fluid-010518-040547 2019

  5. [5]

    and Noack, Bernd R

    Brunton, Steven L. and Noack, Bernd R. and Koumoutsakos, Petros. Machine Learning for Fluid Mechanics. Annual Review of Fluid Mechanics. 2020. doi:https://doi.org/10.1146/annurev-fluid-010719-060214

    doi:10.1146/annurev-fluid-010719-060214 2020

  6. [6]

    Deep neural networks for data-driven LES closure models , journal =

    Andrea Beck and David Flad and Claus-Dieter Munz , keywords =. Deep neural networks for data-driven LES closure models , journal =. 2019 , issn =. doi:https://doi.org/10.1016/j.jcp.2019.108910 , url =

    arXiv 2019

  7. [7]

    and San, O

    Maulik, R. and San, O. and Rasheed, A. and Vedula, P. , year=. Subgrid modelling for two-dimensional turbulence using neural networks , volume=. doi:10.1017/jfm.2018.770 , journal=

    doi:10.1017/jfm.2018.770 2018

  8. [8]

    Toward neural-network-based large eddy simulation: application to turbulent channel flow , volume=

    Park, Jonghwan and Choi, Haecheon , year=. Toward neural-network-based large eddy simulation: application to turbulent channel flow , volume=. doi:10.1017/jfm.2020.931 , journal=

    doi:10.1017/jfm.2020.931 2020

  9. [9]

    and van Leeuwen, C

    Stoffer, R. and van Leeuwen, C. M. and Podareanu, D. and Codreanu, V. and Veerman, M. A. and Janssens, M. and Hartogensis, O. K. and van Heerwaarden, C. C. , TITLE =. Geoscientific Model Development , VOLUME =. 2021 , NUMBER =

    2021

  10. [10]

    Physical invariance in neural networks for subgrid-scale scalar flux modeling , author =. Phys. Rev. Fluids , volume =. 2021 , month =. doi:10.1103/PhysRevFluids.6.024607 , url =

    work page doi:10.1103/physrevfluids.6.024607 2021

  11. [11]

    Jansen and John A

    Aviral Prakash and Kenneth E. Jansen and John A. Evans , keywords =. Invariant data-driven subgrid stress modeling in the strain-rate eigenframe for large eddy simulation , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.cma.2022.115457 , url =

    arXiv 2022

  12. [12]

    Jansen and John A

    Aviral Prakash and Kenneth E. Jansen and John A. Evans , keywords =. Invariant data-driven subgrid stress modeling on anisotropic grids for large eddy simulation , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.cma.2024.116807 , url =

    arXiv 2024

  13. [13]

    Neural-network-based mixed subgrid-scale model for turbulent flow , volume=

    Kang, Myeongseok and Jeon, Youngmin and You, Donghyun , year=. Neural-network-based mixed subgrid-scale model for turbulent flow , volume=. doi:10.1017/jfm.2023.260 , journal=

    work page doi:10.1017/jfm.2023.260 2023

  14. [14]

    A recursive neural-network-based subgrid-scale model for large eddy simulation: application to homogeneous isotropic turbulence , volume=

    Cho, Chonghyuk and Park, Jonghwan and Choi, Haecheon , year=. A recursive neural-network-based subgrid-scale model for large eddy simulation: application to homogeneous isotropic turbulence , volume=. doi:10.1017/jfm.2024.992 , journal=

    work page doi:10.1017/jfm.2024.992 2024

  15. [15]

    MacArt and Jonathan B

    Justin Sirignano and Jonathan F. MacArt and Jonathan B. Freund , keywords =. DPM: A deep learning PDE augmentation method with application to large-eddy simulation , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.jcp.2020.109811 , url =

    arXiv 2020

  16. [16]

    Proceedings of The 33rd International Conference on Machine Learning , pages =

    Group Equivariant Convolutional Networks , author =. Proceedings of The 33rd International Conference on Machine Learning , pages =. 2016 , editor =

    2016

  17. [17]

    Weiler, Maurice and Cesa, Gabriele , booktitle=

  18. [18]

    and Bruna, Joan and Cohen, Taco and Veli

    Bronstein, Michael M. and Bruna, Joan and Cohen, Taco and Veli. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges , journal =. 2021 , url =

    2021

  19. [19]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    E(n) Equivariant Graph Neural Networks , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  20. [20]

    and Kornbluth, Mordechai and Molinari, Nicola and Smidt, Tess E

    Batzner, Simon and Musaelian, Albert and Sun, Lixin and Geiger, Mario and Mailoa, Jonathan P. and Kornbluth, Mordechai and Molinari, Nicola and Smidt, Tess E. and Kozinsky, Boris , title =. Nature Communications , year =. doi:10.1038/s41467-022-29939-5 , url =

    doi:10.1038/s41467-022-29939-5

  21. [21]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Equivariance versus Augmentation for Spherical Images , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    2022

  22. [22]

    Equivariant Networks: A Theory of Generalization on Dynamics Forecasting , author=

    Data Augmentation vs. Equivariant Networks: A Theory of Generalization on Dynamics Forecasting , author=. 2022 , eprint=

    2022

  23. [23]

    2024 , eprint=

    Optimization Dynamics of Equivariant and Augmented Neural Networks , author=. 2024 , eprint=

    2024

  24. [24]

    2024 , eprint=

    Emergent Equivariance in Deep Ensembles , author=. 2024 , eprint=

    2024

  25. [25]

    2025 , eprint=

    Ensembles provably learn equivariance through data augmentation , author=. 2025 , eprint=

    2025

  26. [26]

    Discretize first, filter next: Learning divergence-consistent closure models for large-eddy simulation , journal =

    Syver Døving Agdestein and Benjamin Sanderse , keywords =. Discretize first, filter next: Learning divergence-consistent closure models for large-eddy simulation , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.jcp.2024.113577 , url =

    arXiv 2025

  27. [27]

    2026 , eprint=

    Comparison of data-driven symmetry-preserving closure models for large-eddy simulation , author=. 2026 , eprint=

    2026

  28. [28]

    2026 , eprint=

    Turbulence teaches equivariance to neural networks , author=. 2026 , eprint=

    2026

  29. [29]

    and McConkey, Ryley , title =

    Balla, Julia and Bailey, Jeremiah and Backour, Ali and Hofgard, Elyssa and Jaakkola, Tommi and Smidt, Tess E. and McConkey, Ryley , title =. arXiv preprint arXiv:2509.20683 , note =. 2025 , url =

    arXiv 2025

  30. [30]

    , title =

    Pope, Stephen B. , title =. 2000 , address =

    2000

  31. [31]

    2006 , edition =

    Sagaut, Pierre , title =. 2006 , edition =

    2006

  32. [32]

    , title =

    Germano, Massimo and Piomelli, Ugo and Moin, Parviz and Cabot, William H. , title =. Physics of Fluids A: Fluid Dynamics , volume =. 1991 , doi =

    1991

  33. [33]

    Annual Review of Fluid Mechanics , volume =

    Meneveau, Charles and Katz, Joseph , title =. Annual Review of Fluid Mechanics , volume =. 2000 , doi =

    2000

  34. [34]

    and Haering, Sigfried W

    Moser, Robert D. and Haering, Sigfried W. and Yalla, Gopal R. , title =. Annual Review of Fluid Mechanics , volume =. 2021 , doi =

    2021

  35. [35]

    Cascades in wall-bounded turbulence , journal =

    Jim\'. Cascades in wall-bounded turbulence , journal =. 2012 , doi =

    2012

  36. [36]

    Coherent structures in wall-bounded turbulence , journal =

    Jim\'. Coherent structures in wall-bounded turbulence , journal =. 2018 , doi =

    2018

  37. [37]

    and Mathis, R

    Marusic, I. and Mathis, R. and Hutchins, N. , title =. Science , volume =. 2010 , doi =

    2010

  38. [38]

    and Park, George Ilhwan , title =

    Bose, Sanjeeb T. and Park, George Ilhwan , title =. Annual Review of Fluid Mechanics , volume =. 2018 , doi =

    2018

  39. [39]

    Advances in Neural Information Processing Systems , volume =

    Um, Kiwon and Brand, Robert and Fei, Yun (Raymond) and Holl, Philipp and Thuerey, Nils , title =. Advances in Neural Information Processing Systems , volume =

  40. [40]

    and Alieva, Ayya and Wang, Qing and Brenner, Michael P

    Kochkov, Dmitrii and Smith, Jamie A. and Alieva, Ayya and Wang, Qing and Brenner, Michael P. and Hoyer, Stephan , title =. Proceedings of the National Academy of Sciences , volume =. 2021 , doi =

    2021

  41. [41]

    Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons , journal =

    List, Bj\". Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons , journal =. 2022 , doi =

    2022

  42. [42]

    and Sirignano, Justin and Freund, Jonathan B

    MacArt, Jonathan F. and Sirignano, Justin and Freund, Jonathan B. , title =. Physical Review Fluids , volume =. 2021 , doi =

    2021

  43. [43]

    , title =

    Sanderse, Benjamin and Stinis, Panos and Maulik, Romit and Ahmed, Shady E. , title =. Foundations of Data Science , volume =. 2025 , doi =

    2025

  44. [44]

    Journal of Advances in Modeling Earth Systems , volume =

    Bolton, Thomas and Zanna, Laure , title =. Journal of Advances in Modeling Earth Systems , volume =. 2019 , doi =

    2019

  45. [45]

    Journal of Advances in Modeling Earth Systems , volume =

    Perezhogin, Pavel and Zhang, Cheng and Adcroft, Alistair and Fernandez-Granda, Carlos and Zanna, Laure , title =. Journal of Advances in Modeling Earth Systems , volume =. 2024 , doi =

    2024

  46. [46]

    Geophysical Research Letters , volume =

    Perezhogin, Pavel and Adcroft, Alistair and Zanna, Laure , title =. Geophysical Research Letters , volume =. 2025 , doi =

    2025

  47. [47]

    , title =

    Yuval, Janni and O'Gorman, Paul A. , title =. Nature Communications , volume =. 2020 , doi =

    2020

  48. [48]

    Journal of Computational Physics , volume =

    Guan, Yifei and Chattopadhyay, Ashesh and Subel, Adam and Hassanzadeh, Pedram , title =. Journal of Computational Physics , volume =. 2022 , doi =

    2022

  49. [49]

    Physica D: Nonlinear Phenomena , volume =

    Guan, Yifei and Subel, Adam and Chattopadhyay, Ashesh and Hassanzadeh, Pedram , title =. Physica D: Nonlinear Phenomena , volume =. 2023 , doi =

    2023

  50. [50]

    , title =

    Speziale, Charles G. , title =. Journal of Fluid Mechanics , volume =. 1985 , doi =

    1985

  51. [51]

    Journal of Turbulence , volume =

    Li, Yi and Perlman, Eric and Wan, Minping and Yang, Yunke and Meneveau, Charles and Burns, Randal and Chen, Shiyi and Szalay, Alexander and Eyink, Gregory , title =. Journal of Turbulence , volume =. 2008 , doi =

    2008

  52. [52]

    and Kanov, K

    Graham, J. and Kanov, K. and Yang, X. I. A. and Lee, M. and Malaya, N. and Lalescu, C. C. and Burns, R. and Eyink, G. and Szalay, A. and Moser, R. D. and Meneveau, C. , title =. Journal of Turbulence , volume =. 2016 , doi =

    2016

  53. [53]

    , title =

    Lee, Myoungkyu and Moser, Robert D. , title =. Journal of Fluid Mechanics , volume =. 2015 , doi =

    2015

  54. [54]

    International Conference on Learning Representations , year =

    Cesa, Gabriele and Lang, Leon and Weiler, Maurice , title =. International Conference on Learning Representations , year =

  55. [55]

    and Ba, Jimmy , title =

    Kingma, Diederik P. and Ba, Jimmy , title =. International Conference on Learning Representations , year =

  56. [56]

    Advances in Neural Information Processing Systems , volume =

    Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and K\". Advances in Neural Information Processing Systems , volume =

  57. [57]

    2026 , howpublished =

    McConkey, Ryley and Balla, Julia , title =. 2026 , howpublished =

    2026

  58. [58]

    Banerjee and R

    S. Banerjee and R. Krahl and F. Durst and Ch. Zenger , title =. Journal of Turbulence , volume =. 2007 , publisher =. doi:10.1080/14685240701506896 , URL =

    work page doi:10.1080/14685240701506896 2007

  59. [59]

    2025 , eprint=

    Does equivariance matter at scale? , author=. 2025 , eprint=

    2025