{"id":"7cd15b9a-4187-4554-866d-39fa6218045c","arxiv_id":"2607.26850","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"At realistic LES filter ratios on channel flow, an octahedral-equivariant nonlocal CNN is more accurate, parameter-efficient, and data-efficient than an augmented non-equivariant CNN, while pointwise models fail to beat the Clark baseline.","lead":"Equivariant nonlocal neural nets beat data-augmented non-equivariant ones for predicting unresolved turbulence stresses, at roughly half the parameters. The result gives practical design guidance for machine-learned closures used in large-eddy simulation.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the a-priori/a-posteriori gap already flagged; the controlled bake-off supports the stated claim.","rationale":"The central claim is a controlled a-priori bake-off, not an unqualified deployment recommendation. Evidence quality for that narrower claim is adequate: public data, clear spatial/temporal splits, analytical Clark baseline, matched-parameter sweeps, and consistent ESCNN superiority across spatiotemporal, anisotropy, and Reynolds axes (Table 3). The a-priori/a-posteriori gap is real and already correctly identified by the reader as the reason for CONDITIONAL rather than unconditional ACCEPT; the paper itself flags it in §4. Secondary issues (single seed, steerable-conv implementation cost) affect confidence intervals and practical wall-clock but do not reverse the reported ranking inside the stated scope. Therefore no verdict adjustment is warranted; multi-seed confirmation remains the natural next verification step before any stronger claim.","tokens_in":24797,"tokens_out":500,"duration_ms":9160,"concrete_test":"Re-train the Table-3 ESCNN (46.9k) and augmented CNN (95.4k) configurations with three independent seeds on the same near-wall Re_τ=1000 split; recompute ρ_τ on the three generalization sets. If ESCNN remains strictly highest on every cell at ~half the parameters, the a-priori claim is robust to seed variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader already isolates the structurally weakest premise: that a-priori ρ_τ rankings on box-filtered DNS (filter ratio 4, Δ+≈40, velocity-gradient inputs on 14^{3} interiors) transfer to LES deployment where stability and dissipation matter. The paper states this limitation explicitly (§4) and never claims in-solver results. Within the a-priori scope actually argued, the matched-parameter comparison (Table 3; Figs. 5–6), Clark baseline, three generalization axes, and public JHTDB extraction are internally consistent. Single-seed training and research-code wall-clock for steerable convolutions are secondary and do not overturn the ranking of ESCNN over the augmented CNN at practical parameter counts. No additional load-bearing internal inconsistency was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","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.","tokens_in":24997,"tokens_out":1372,"duration_ms":42250,"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":[{"comment":"§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.","section":"§4 Conclusion; Abstract; Table 3"},{"comment":"§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.","section":"§2.5 Training procedure; Table 3; §3.2–3.3"}],"minor_comments":[{"comment":"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.","section":"§3.2; Figures 5–6"},{"comment":"§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.","section":"§2.2"},{"comment":"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).","section":"Table 1; §3.3"},{"comment":"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.","section":"Abstract; throughout"},{"comment":"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.","section":"§2.3.3; Appendix A"},{"comment":"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.","section":"§1 Introduction"}],"recommendation":"minor_revision","confidential_remarks":"Fit for a fluids/computational-mechanics journal is good. The a priori limitation is standard for this literature and is disclosed; I would not block on full a posteriori LES if the conclusion is tightened. No concerns about circularity or data provenance (JHTDB). Self-citations to the authors’ super-resolution/implicit-equivariance preprints are substantive and relevant, not excessive. Single-seed and research-code timing are the main polish items."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing: at realistic filter ratio on channel flow, a steerable octahedral CNN beats a data-augmented plain CNN on every a-priori generalization axis they tried, at roughly half the parameter count, and pointwise nets (equivariant eigenframe or not) sit on top of Clark. That is the paper.\n\nWhat is actually new is not the architectures—Prakash eigenframe, steerable CNNs, octahedral aug, nonlocal SGS CNNs are all prior—but the controlled matched-parameter cross of equivariance × locality on SGS stress, plus a quantitative look at implicit octahedral equivariance learned from turbulence alone (extending their super-resolution result). They did the comparison carefully: JHTDB boxes, clear space/time splits, Δ+≈40, velocity-gradient-only Galilean-invariant inputs, Δ²G² nondimensionalization so Re transfer is not a rescaling trick, Clark baseline, and three generalization axes (spatiotemporal, anisotropy, Re). Table 3 and the size/data sweeps are the core evidence and they hang together. Code is promised on GitHub.\n\nSoft spots, in proportion. Single seed and no uncertainty bands—annoying for a ranking paper, not fatal given the consistency across sweeps. Wall-clock favors the plain CNN partly because steerable convs are research code; they say so. The real limit is the one they state in §4: everything is a-priori ρ_τ on box-filtered DNS. No LES insertion, no stability or dissipation check. So the practical “use this in solvers” reading is a transfer assumption, not a result. Within the a-priori scope they actually argue, I do not see an internal crack.\n\nThis is for people building data-driven SGS or deciding whether to pay the equivariant-architecture tax in fluids ML. Not a theory paper. I would bring it to reading group, cite it when discussing inductive bias for closures, and send it to referees. Accept-shaped empirical contribution if the claims stay a priori; any deployment language needs the usual a-posteriori caveat.","headline":"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.","tokens_in":25675,"tokens_out":539,"would_cite":true,"duration_ms":14761,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Equivariant nonlocal networks beat data-augmented ones for subgrid stress at half the parameters, while pointwise models barely beat Clark.","keywords":["large eddy simulation","subgrid-scale modelling","data-driven turbulence closure","equivariant neural networks","rotational octahedral group","nonlocality","channel flow","parameter efficiency"],"falsifier":"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.","tokens_in":25605,"feed_emoji":"🌀","tokens_out":988,"duration_ms":16739,"temperature":0.7,"pith_summary":"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.","feed_headline":"Equivariant nonlocal nets win SGS stress at half the parameters","feed_subtitle":"Pointwise models fail to beat Clark; both equivariance and nonlocality matter at realistic filter widths","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Equivariant nonlocal SGS nets top every test at half the parameters","Nonlocal equivariance beats data-augmented CNNs on channel-flow SGS","Pointwise models stall at Clark; equivariance gains grow with receptive field","Equivariant nonlocal closures win correlation and data efficiency","Rotational equivariance plus nonlocality cut parameters for SGS stress"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant nonlocal SGS nets top every test at half the parameters","Nonlocal equivariance beats data-augmented CNNs on channel-flow SGS","Pointwise models stall at Clark; equivariance gains grow with receptive field","Equivariant nonlocal closures win correlation and data efficiency","Rotational equivariance plus nonlocality cut parameters for SGS stress"]},"model":"grok-4.5","effort":"low","cost_usd":0.001813,"raw_usage":{"total_tokens":924,"prompt_tokens":824,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":18128000,"prompt_tokens_details":{"text_tokens":824,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":824,"tokens_out":80,"duration_ms":2630,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T19:13:54.916370+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"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.","supporting_citations":[],"review_version":1}