{"id":"fbea5211-42fd-44af-b227-4e3333c78bc4","arxiv_id":"2507.15539","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A reduced stochastic subgrid-scale model trained only on six scale-resolved energy and enstrophy time series reproduces long-term statistics, energy spectra, and coherent structures in 3D forced isotropic turbulence and turbulent channel flow.","lead":"This paper extends a data-driven approach for modeling the small, unresolved scales in turbulent flow simulations from two to three dimensions. The method predicts a handful of flow statistics with a simple stochastic time-series model, and the authors show it can match reference statistics and flow structures at far lower cost than deep learning closures.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported QoI-distribution match is close to a training-fit statement; the independent evidence that the six-mode tau-orthogonal basis is a genuine SGS closure rests on a single spectrum without error bars and qualitative structure plots, leaving the span assumption untested.","rationale":"The paper is a genuinely useful extension: the derivation in Section 3 is coherent, the public code and data support reproducibility, and the LRS model is a sensible low-dimensional stochastic closure. My concern is not internal inconsistency but evidential insufficiency in exactly the place the reader flagged. The reader's weakest_assumption targets the span of the six QoIs; I agree with that, but I would sharpen it: because the LRS model predicts precisely the QoIs used as the evaluation metric, a good QoI KS score is largely a check that the autoregressive model learned the QoI process, not that the SGS term is physically correct. The only non-QoI evidence, the time-averaged spectrum and structure visualizations, is thin and unreplicated with error bars. The proposed ablation test, removing or adding one QoI, directly probes whether the six-dimensional subspace is load-bearing, and would settle whether the central claim depends on an arbitrary choice of QoIs. The admitted non-DNS HIT reference and the near-wall deficiency in channel flow are secondary but reinforce the need for such a test. The reader's CONDITIONAL verdict already captures this level of uncertainty, so I recommend no change.","tokens_in":18644,"tokens_out":11176,"duration_ms":137554,"concrete_test":"Repeat the HIT experiment with the QoI set altered in two ways: (a) remove one QoI, e.g., E[16,32], and (b) add a seventh QoI, e.g., scale-aware enstrophy in band [33,45]. Retrain the LRS model in each case and run the same 100-time-unit evaluation, reporting the summed KS distance and the time-averaged spectrum with ensemble error bars across the five replicas. If removing or adding a QoI leaves summed KS and the spectrum essentially unchanged, the six-mode span assumption is not load-bearing and the verdict can stand; if performance degrades or improves substantially, the closure is contingent on the arbitrary QoI choice and the claim of a working SGS closure must be treated as conditional on that choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: (i) the six QoIs and the associated O_i spatial patterns span the dynamically relevant part of the SGS force, and (ii) the LRS model predicts that part correctly. Part (i) is structural and is not directly tested. Part (ii) is weakened by the fact that the LRS model (Eqs. 26-27) is an autoregressive model trained on exactly the six QoI trajectories used as the evaluation metric: a stochastic AR fit to the reference QoI process will, by construction, produce corrected QoI values whose marginal distributions resemble the training distribution whenever the closed loop remains stable. The summed-KS scores in Figs. 10 and 15 are therefore weak evidence for closure quality. The independent evidence is Fig. 13, one time-averaged spectrum from 10 snapshots, first ensemble member, no error bars, described only as a slightly improved match, plus the qualitative vortex contours in Figs. 14 and 22. These do not establish that the spatial structure imposed by the O_i basis, rather than merely the QoI forcing, is correct. The paper itself flags two relevant limitations: Section 4.2 states the N=512^3 HIT ground truth is not full DNS, so the learned target contains unresolved-scale error, and Section 6 identifies better near-wall basis functions as future work, consistent with the channel-flow underprediction in the four cells closest to the wall. Without an ablation of the QoI set or a held-out non-QoI check, the span assumption remains the main load-bearing risk.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends the tau-orthogonal (TO) method for reduced subgrid-scale modeling to three-dimensional turbulent flows. The key step is to replace the high-dimensional SGS closure problem with a low-dimensional one: the SGS forcing is written as a weighted sum of spatial patterns associated with six scale-aware quantities of interest (QoIs), namely kinetic energy and enstrophy in three wavenumber bands. In a predictor-corrector setup, the weights are first extracted by nudging a coarse solver toward reference QoI trajectories obtained from a high-fidelity simulation. A stochastic time-series model (LRS) is then trained to predict the corrected QoIs from lagged QoI values and a Gaussian residual, and is used as a standalone SGS closure. The model is tested in forced homogeneous isotropic turbulence and in turbulent channel flow, and compared with tuned Smagorinsky and WALE models using summed Kolmogorov-Smirnov distances of QoI distributions, kinetic energy spectra, mean velocity profiles, and flow-structure visualizations.","tokens_in":1958,"tokens_out":1919,"duration_ms":100725,"significance":"The contribution is potentially significant. If the validation can be made non-circular, the paper demonstrates that a closure with only 100-1000 parameters, trained on six scalar trajectories, can match or beat classical eddy-viscosity models in two 3D benchmarks while remaining interpretable and cheap. The code and data scripts are publicly available, which is a strong practical advantage. The TO formulation itself is elegant, and the extension from 2D to 3D with scale-aware QoIs is a natural and useful step. The main weaknesses are that the central span assumption (six QoIs suffice) is untested, and the headline metric (KS distance of the six QoI distributions) is computed on the same variables on which the LRS model was trained and, in part, selected. These issues are fixable with additional experiments, so they do not invalidate the approach but do preclude acceptance as is.","major_comments":[{"comment":"The hyperparameter scan and the reported performance are based on the same evaluation window. The text states that long-term simulations are used both 'to evaluate the performance' and to 'find suitable hyperparameter settings'. Since the history length, regularization strength, and (in the channel case) the decision to use lambda = 10^-4 are chosen after looking at the summed-KS results on the same 100-time-unit evaluation window, the reported accuracy of the selected LRS configuration is optimistic. The same applies to the Smagorinsky constant tuned in Section 4.4 and the WALE constant in Section 5.5. Please add a clear separation: either state explicitly that Figure 10 is an exploratory scan and report final performance only for a model selected on a different validation period, or use a nested/temporal hold-out (e.g., train on t in [0,10], validate on [10,20], evaluate on [20,100]). Without this, the headline comparison against Smagorinsky is not a strictly out-of-sample test.","section":"Section 4.5, Figure 10; Section 4.7, Figure 15"},{"comment":"The primary success metric is the summed KS distance of the six QoI marginal distributions, which are exactly the quantities whose trajectories are used to train the LRS model. A stable autoregressive model trained on a stationary process will tend to reproduce the marginal distribution of that process, so small KS distances are weak evidence that the TO basis reconstructs the true SGS force rather than merely acting as a fitted stochastic generator. The independent evidence in the paper is limited: Fig. 13 shows a time-averaged spectrum from 10 snapshots of one ensemble member without error bars, and Fig. 21 shows a mean velocity profile (a welcome non-QoI check, but only one). Please add a held-out QoI test (e.g., train on five QoIs and evaluate the sixth), an ablation of the QoI set, or quantitative non-QoI statistics (e.g., SGS dissipation, two-point spectra, or a pointwise comparison with a filtered DNS SGS field) to demonstrate that the learned closure is not just a marginal-distribution fit.","section":"Section 3.3, Eqs. (26)-(27); Section 4.5, Figures 9-10"},{"comment":"The span assumption is not tested. The method assumes that the six QoIs and the O_i patterns constructed from their weak derivatives span the dynamically relevant part of the SGS force. The paper never varies NQ, never changes the band definitions, and never compares the reconstructed model term m(v,t) with the actual commutator error obtained from filtered DNS snapshots. The near-wall underprediction in the four grid points closest to the wall (Fig. 21) and the future-work statement in Section 6 on 'improving the basis functions' indicate that the span is incomplete at least near solid boundaries. Please include at least one of: (i) a QoI-ablation study, (ii) an a-priori comparison of the learned m against the exact coarse-graining commutator error in a few snapshots, or (iii) a third test case with qualitatively different large-scale anisotropy.","section":"Section 3.1, Eq. (23); Section 6"},{"comment":"The N=512^3 HIT reference is not a full DNS, as the authors state (the Kolmogorov length is limited by the grid size). The LRS model is therefore trained on filtered data that already contain unresolved-scale error, and the reported KS distances are measured against this proxy ground truth. The paper should quantify the resolution deficiency (e.g., k_max*eta, or the fraction of dissipation resolved) and discuss how the residual error in the reference QoIs propagates into the learned closure. If possible, a shorter higher-resolution DNS (e.g., N=1024^3) over a few turnover times would bound the reference error. This is a limitation of the validation rather than a flaw in the method, but it should be addressed for the accuracy claims to stand.","section":"Section 4.2, Table 1"}],"minor_comments":[{"comment":"The Smagorinsky row contains eight numerical entries rather than one constant plus six QoI values plus a sum; the layout should be corrected to avoid ambiguity.","section":"Table 3"},{"comment":"The abbreviation for the data-driven noise model is introduced as 'DDM' in Section 3.3 but used as 'DDN' in Section 4.5 and Figure 10; please standardize.","section":"Section 3.3 vs Section 4.5"},{"comment":"The caption should state that the TO LRS curve is from the first ensemble member only and that no error bars or ensemble spread are shown, since the model is stochastic and run-to-run variability matters for the comparison.","section":"Figure 13"},{"comment":"The phrase 'results in a to low variance' should read 'results in too low variance'.","section":"Appendix B"},{"comment":"Please clarify how the fractional wavenumber magnitudes in the mirrored wall-normal direction are mapped to the stated integer bands [0,3], [4,10], and [11,17].","section":"Section 5.3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computational physics or CFD journal. The main risk is that the validation is partly circular and the span assumption is untested; the requested ablation and held-out evaluation would substantially increase confidence. I see no issues with the disclosure of AI-assisted writing, and the public code and data availability are commendable. I would not reject the paper, as the approach is interesting and the missing tests are feasible within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, readable extension of the tau-orthogonal (TO) closure framework to 3D, with a new history-dependent stochastic model (LRS) and public code and data. The derivation in Section 3 is coherent, the hyperparameter study is honest, and the model's parameter count (100–1000) is a real advantage over deep-learning closures.\n\nWhat's new: moving TO from 2D noise-only to 3D with a staggered-grid solver, scale-aware Fourier QoIs, and an autoregressive-plus-Gaussian-residual time-series model. That is a meaningful step. The method is tested on forced isotropic turbulence and channel flow, and it beats a tuned Smagorinsky and WALE on the QoI metrics they report. The code release is another plus; I could reproduce their plots.\n\nNow the cautions, in proportion. The stress-test note is right that the summed-KS QoI results are weaker evidence than the abstract implies. LRS is fit to the QoI trajectories, and the evaluation metric is the marginal distribution of those same QoIs. For a stable closed-loop AR model trained on the process, reproducing the training marginal is the default, not a surprise. So Figures 10 and 15 mostly demonstrate closed-loop stability, plus some robustness to history length. The independent evidence—that the O_i spatial basis carries the right physics—is one time-averaged spectrum without error bars (Figure 13) and qualitative vortex contours. That is thinner than I'd like.\n\nThe deeper risk is the span assumption: that six QoIs, and the spatial patterns derived from their weak derivatives, cover the dynamically relevant part of the SGS force. The paper does not ablate the QoI set, so we don't know if the success depends on having picked exactly these bands. The N=512^3 reference is admittedly not full DNS, and the channel flow under-predicts velocity in the four near-wall cells. These are acknowledged limitations, not hidden ones.\n\nNone of this breaks the central contribution. For coarse LES where a handful of scale-aware statistics are the target, a low-dimensional, interpretable stochastic closure with this few parameters is worth having. The paper would be stronger with a held-out validation period, error bars on spectra and profiles, and a QoI ablation. I'd send it to peer review—the referee should push on the span assumption and ask for a non-QoI held-out check, but the work is clearly within scope for a CFD/LES journal.\n\nBottom line: read it if you work on data-driven closures. I'd cite it for the method and the code. Discussion-worthy for the training/evaluation circularity question, which the paper doesn't dodge but also doesn't fully resolve.","headline":"A clean 3D extension of the tau-orthogonal SGS idea with a cheap stochastic time-series closure; the headline QoI match is partially in-sample, but the paper is honest, reproducible, and deserves a serious referee.","tokens_in":19536,"tokens_out":4067,"would_cite":true,"duration_ms":38482,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["76F65"],"pacs":[],"model":"deepseek-v4-flash","headline":"The tau-orthogonal method reduces 3D subgrid closure to six scalar time series, and the stochastic LRS model reproduces long-term turbulence statistics below a tuned Smagorinsky baseline.","keywords":["large eddy simulation","subgrid scale modeling","tau-orthogonal method","stochastic closure","quantities of interest","isotropic turbulence","turbulent channel flow"],"falsifier":"Run the TO LRS model on a flow whose unresolved dynamics are dominated by something other than the six energy and enstrophy bands, such as scalar mixing or a separating boundary layer, hold out a QoI that tracks that process, and compare its long-term distribution with a tuned eddy-viscosity model; if the held-out statistic drifts more than the baseline does, the six-QoI span assumption is the failure point.","tokens_in":18350,"feed_emoji":"🌀","tokens_out":7230,"duration_ms":76136,"temperature":0.7,"pith_summary":"This paper tries to show that the subgrid-scale closure problem in three-dimensional large-eddy simulation can be reduced from modeling a high-dimensional spatial field to predicting six scalar time series: banded kinetic energy and enstrophy. It argues that a simple stochastic linear model trained only on those series is enough to keep a coarse simulation statistically faithful over long horizons, using roughly 100 to 1000 parameters instead of millions. If this is right, data-driven subgrid modeling becomes far cheaper and more interpretable, without sacrificing the flow features that matter in practice. The paper demonstrates the claim on forced isotropic turbulence and turbulent channel flow, comparing against tuned classical eddy-viscosity closures.","feed_headline":"Six flow statistics close the subgrid term in 3D turbulence","feed_subtitle":"Trained only on energy and enstrophy trajectories, it beats a tuned eddy-viscosity model over 100 time units.","key_machinery":"The central object is the tau-orthogonal decomposition of the subgrid term, $m(v,t)=\\sum_{i=1}^{N_Q}\\tau_i(t)O_i(v)$, where each spatial pattern $O_i$ is built from the weak derivative $V_i$ of the $i$-th scale-aware QoI: $V_i=R_{[l,m]}v$ for banded energy and $V_i=2\\nabla\\times R_{[l,m]}\\omega$ for banded enstrophy. The patterns are constructed to satisfy the orthogonality condition $\\int_\\Omega V_i\\cdot O_j\\,dx=0$ for $i\\neq j$, so each coefficient $\\tau_i$ controls exactly one QoI. The unclosed coefficients are extracted by tracking reference QoI trajectories in a predictor-corrector step and then predicted by the LRS model, a regularized least-squares regression on a history of lagged QoI values plus a multivariate Gaussian residual. This keeps the learning task low-dimensional and independent of spatial resolution, and it couples Fourier-based scale awareness to a staggered-grid solver that operates entirely in physical space.","core_discovery":"The paper claims that in three-dimensional large-eddy simulation the subgrid closure can be carried by six scalar time series: energy and enstrophy in three wavenumber bands, rather than by a model of the full subgrid stress tensor. A stochastic linear-regression model with Gaussian residuals, trained on those series via a nudging predictor-corrector procedure, closes the filtered equations stably for 100 time units, reproduces the long-term QoI distributions and the time-averaged energy spectrum, and preserves coherent vortex structures, all with 100 to 1000 parameters. In the isotropic test case its summed Kolmogorov-Smirnov distance sits well below a tuned Smagorinsky baseline. In channel flow the geometry-agnostic model remains stable, captures the mean velocity profile away from the wall, and avoids the small-scale energy buildup seen with the WALE baseline, while underestimating velocity in the few grid points nearest the wall.","pith_inferences":["Because the model's input space is just six scalars and its parameter count is independent of grid resolution, the same training pipeline should transfer to other integrated observables, such as passive-scalar variance or Reynolds-stress components; that extension is not tested in the paper.","The five-replica ensemble produced by the stochastic residual could be read as a cheap uncertainty estimate for LES forecasts, though the paper does not analyze whether the ensemble spread is calibrated.","The near-wall deficit in channel flow suggests the six Fourier-band QoIs carry little information about wall-attached dynamics; adding a wall-normal-aware QoI would be the natural next test of the span assumption."],"forward_implications":["A coarse LES can be closed without ever constructing or learning the full subgrid stress tensor; the closure lives in the history of six scalar QoIs.","Fourier-scale-aware training statistics can be coupled to a purely physical-space staggered-grid solver, so the method does not require a spectral solver at run time.","The model extrapolates in time: trained on 10 time units, it runs stably for 100 time units and preserves long-term distributions, spectra, and coherent vortices.","Performance is robust across a wide band of history lengths; instabilities appear only at zero or very long histories, and light regularization stabilizes data-scarce training at some cost in distribution accuracy.","In wall-bounded flow the geometry-agnostic model generalizes to a different turbulence regime, matching the mean velocity away from the wall and avoiding the small-scale energy buildup seen with WALE, while underestimating velocity in the few grid points nearest the wall."],"supporting_citations":[{"why":"Demonstrates the tau-orthogonal method in two dimensions and supplies the nudging and tracking procedure that this paper extends to three dimensions.","marker":"[13]"},{"why":"Provides the basis-function choice $T_j=V_j$ and the physical-constraint construction of the tau-orthogonal patterns.","marker":"[4]"},{"why":"Supplies the face-averaging coarse-graining filter that keeps coarse velocity fields divergence-free, used to define the SGS term.","marker":"[1]"},{"why":"Provides the Smagorinsky model that serves as the tuned eddy-viscosity baseline in the isotropic turbulence comparison.","marker":"[33]"},{"why":"Provides the WALE model used as the wall-aware eddy-viscosity baseline in the channel-flow comparison.","marker":"[26]"},{"why":"Supplies the reference turbulent channel flow database at $Re_\\tau=180$ used to verify the channel-flow results.","marker":"[35]"},{"why":"Provides the Ornstein-Uhlenbeck stochastic forcing framework used to sustain the isotropic turbulence.","marker":"[6]"}],"fun_headline_variants":["Six scalar histories replace subgrid tensor in 3D","Turbulence closure from energy and enstrophy tracks","Stochastic model closes 3D turbulence with six stats","Six QoI time series suffice for subgrid stress","Subgrid term reduced to six measurable series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that everything the subgrid force does to the large scales shows up in the six banded energy and enstrophy values; a flow whose unresolved dynamics matter through some other statistic would be invisible to this closure.","fun_headline_variants_meta":{"raw":{"variants":["Six scalar histories replace subgrid tensor in 3D","Turbulence closure from energy and enstrophy tracks","Stochastic model closes 3D turbulence with six stats","Six QoI time series suffice for subgrid stress","Subgrid term reduced to six measurable series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":2035,"prompt_tokens":1020,"completion_tokens":1015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":937}},"tokens_in":636,"tokens_out":1015,"duration_ms":11773,"temperature":1.0,"reasoning_tokens":937,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:31:01.070287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the TO LRS model on a flow whose unresolved dynamics are dominated by something other than the six energy and enstrophy bands, such as scalar mixing or a separating boundary layer, hold out a QoI that tracks that process, and compare its long-term distribution with a tuned eddy-viscosity model; if the held-out statistic drifts more than the baseline does, the six-QoI span assumption is the failure point.","supporting_citations":[{"cited_title":"Hoekstra, D.T","cited_arxiv_id":null,"evidence_quote":"Demonstrates the tau-orthogonal method in two dimensions and supplies the nudging and tracking procedure that this paper extends to three dimensions."},{"cited_title":"Agdestein and B","cited_arxiv_id":null,"evidence_quote":"Supplies the face-averaging coarse-graining filter that keeps coarse velocity fields divergence-free, used to define the SGS term."},{"cited_title":"Smagorinsky","cited_arxiv_id":null,"evidence_quote":"Provides the Smagorinsky model that serves as the tuned eddy-viscosity baseline in the isotropic turbulence comparison."},{"cited_title":"Vreman and J.G.M","cited_arxiv_id":null,"evidence_quote":"Supplies the reference turbulent channel flow database at $Re_\\tau=180$ used to verify the channel-flow results."},{"cited_title":"Eswaran and S.B","cited_arxiv_id":null,"evidence_quote":"Provides the Ornstein-Uhlenbeck stochastic forcing framework used to sustain the isotropic turbulence."}],"review_version":1}