{"id":"df686474-90b3-4e7c-a2ee-e3f739af9149","arxiv_id":"2504.14473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Rotation-and-reflection-invariant machine learning inputs improve a priori prediction of the signed Smagorinsky coefficient, including backscatter regions, for mesoscale hurricane boundary layer flows.","lead":"This paper trains machine learning models to predict a signed turbulence coefficient that represents both energy flow from large to small eddies and the reverse, backscatter, in hurricane boundary layers. The models are tested against stored simulation data, not yet inside a running hurricane simulation, so the practical payoff is still unproven.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported a priori gains may reflect data leakage: overlapping group patches and test-based ensemble selection are never separated by a strict train/test split.","rationale":"The paper is best read as an a priori ML fitting study, and within that scope its central claim is that an invariance-embedded ML model can predict the signed Smagorinsky coefficient accurately enough to outperform the dynamic Smagorinsky model. For that claim to hold, the reported classification and regression metrics must be unbiased estimates of performance on unseen flow states. The manuscript does not document a train/test split at the level of flow fields: it only says the dataset was 'randomized' (Section IV A 2) and that group learning uses overlapping 3x3/5x5 neighborhoods (Section III C 2). With point-level randomization, a test point's neighborhood overlaps training neighborhoods, so the model has effectively seen the test point's local pattern during training; this alone can inflate F1 and depress MSE. The ensemble regression description in Section IV B 2 compounds the problem: networks are selected by an MSE threshold before aggregation and the resulting 'test MSE' is reported as low as 2.49e-8. If the threshold is computed on the test set, the reported number is the result of selection on the test set, not an estimate of generalization error. Both issues are standard methodological pitfalls, and the paper can be repaired by reporting a strict spatial/SST split and nested model selection. I do not regard the Smagorinsky projection limitation as the primary risk, because the paper's stated deliverable is exactly a signed scalar coefficient and the limitation is explicit in the chosen carrier model. The leakage/selection issue, by contrast, threatens the empirical comparison itself. If the numbers survive a strict re-run, the central claim stands; if they do not, the 'superior performance' assertion is unsupported. The lack of released code and the absence of error bars further hinder independent verification, but those are secondary to the need for a correctly structured evaluation.","tokens_in":20790,"tokens_out":7236,"duration_ms":68818,"concrete_test":"Re-run the full classification and regression pipeline with a strict generalization split: train on the 26, 27, and 28 degrees C SST simulations and test on 29 degrees C (or, if SST mixing is intended, split the 62 m domain into disjoint spatial blocks with a buffer of at least one filter width), and select ensemble members only by out-of-bag or validation MSE, never by the final test MSE. Compare the resulting F1 and MSE with Tables V/VIII and with DSM on the identical held-out points. If the F1 gain over non-geometric inputs shrinks below about 5 points or the ensemble test MSE rises by more than 10x, the invariance/ensemble superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative claims that support the paper's central assertion, namely the ~17% F1 improvement from geometric invariance and the ensemble NN's 'superior performance' over the dynamic Smagorinsky model, depend on an evaluation protocol that does not rule out data leakage. Section III C 2 constructs each training sample from overlapping 3x3 or 5x5 neighborhoods; if train and test points are randomized at the point level, adjacent test points share nearly identical input patches with training points, making the reported F1 and MSE optimistic measures of interpolation rather than generalization. No spatial, temporal, or SST-based split is described. Additionally, Section IV B 2 states that the ensemble aggregates only networks meeting an MSE threshold of 1e-8 before reporting a final ensemble test MSE as low as 2.49e-8; if that threshold is evaluated on the same test set used for the final number, the reported MSE is a selection statistic, not an unbiased estimate of predictive error. Both issues directly affect the central empirical comparison against DSM. The Smagorinsky alignment assumption identified by the Reader is a genuine limitation of the chosen carrier model, but it is secondary here: it restricts what the scalar can represent, whereas the leakage/selection issue would invalidate the numerical evidence itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage machine-learning sub-grid-scale (SGS) stress closure for meso-scale hurricane boundary layer LES. A binary classifier separates regions of strong signed Smagorinsky coefficient magnitude from weak ones using tensor invariants and local singular-value-decomposition features, and a regression network predicts the signed Smagorinsky coefficient Cs in strong-transfer regions. The study uses high-resolution WRF LES hurricane data, compares pointwise versus group learning, and reports that geometric invariance embedded via 3x3 and 5x5 SVD neighborhoods improves classification F1 by roughly 17%, with ensemble neural networks yielding lower mean squared error than single networks and matching the dynamic Smagorinsky model more closely in a priori tests. The paper positions the work as an interpretable, invariance-respecting alternative to black-box SGS closures while retaining the Smagorinsky functional form for ease of CFD implementation.","tokens_in":21015,"tokens_out":3320,"duration_ms":34012,"significance":"If the reported gains are robust, the framework is a useful contribution: it gives a physically motivated, feature-based route to signed-Cs prediction that explicitly admits energy backscatter, and the SVD-based geometric invariance is a sensible way to make local SGS models coordinate-aware without losing tensorial consistency. The use of production-based Cs as a well-defined scalar target, the systematic comparison across pointwise and group learning, and the ensemble bagging strategy are all strengths. The central empirical claims, however, rest on an evaluation protocol that is not yet demonstrated to be leakage-free: overlapping patches, point-level randomization, and test-set-based ensemble selection could inflate both the F1 improvement and the MSE advantage over DSM. The Smagorinsky alignment assumption is also a genuine modeling limit that should be stated and tested. These issues are fixable within the manuscript's scope, so the result is promising but not yet established.","major_comments":[{"comment":"The train/test protocol is not described in a way that rules out spatial leakage. Group learning constructs each sample from overlapping 3x3 or 5x5 neighborhoods, and the text states only that the dataset was 'randomized and converted to PyTorch tensors'; if points are split at random, adjacent test points share nearly all of their input patches with training points, so the reported F1 values (Tables IV-VI) measure interpolation within a dense patch cloud rather than generalization to unseen flow regions. Please specify whether the split is spatial, temporal, or by sea-surface-temperature case, and report metrics for a spatially separated split, for example by holding out entire eyewall sectors or SST cases.","section":"§III C 2 and §IV A 2"},{"comment":"The ensemble evaluation in Table VIII is vulnerable to selection bias. The text says predictions are aggregated from 'networks meeting a specified MSE threshold (1e-8)' and then reports a final ensemble test MSE as low as 2.49e-8; if that threshold is applied to the same test set used to compute the final number, the reported MSE is a selection statistic rather than an unbiased estimate of predictive error. Please clarify on which data the threshold is evaluated, and use a nested or validation-based selection protocol so that the reported test MSE is not optimized over.","section":"§IV B 2"},{"comment":"The scalar Cs is defined by projecting the true SGS production onto the Smagorinsky strain-rate alignment, and this same projected Cs is simultaneously the training label, the regression output, and the benchmark against which DSM is compared. As a result, the paper demonstrates prediction of a projected scalar, not of the full anisotropic SGS stress tensor; if the real stress has components orthogonal to the deviatoric strain rate, both the ML predictions and the DSM comparison are blind to them. The authors should state this limitation explicitly and, ideally, add a direct evaluation of predicted SGS stress components or production against the filtered LES truth to show that the scalar projection captures the physically relevant part of the stress.","section":"§III B, Eq. (4)"},{"comment":"No error bars, repeated-seed statistics, or cross-validation results are reported for any of the classification or regression metrics. The central claim of a roughly 17% F1 improvement from geometric invariance is a comparison of single numbers; without variance over network initializations and data splits, it is not possible to tell whether that improvement is statistically significant. Please report means and standard deviations over at least several random seeds and over the chosen split, and state the sample sizes used for training and testing in each configuration.","section":"Tables III-VI and VIII"}],"minor_comments":[{"comment":"Equation (4) appears to rely on an incompressibility or commutation assumption for the filtered field, but the manuscript does not state this explicitly; please clarify the validity of S*_ii = 0 for the filtered hurricane boundary layer data.","section":"§III B, Eq. (4)"},{"comment":"The in-text citation 'Lily et al.' should be 'Lilly', and the reference should specify D. K. Lilly's 1967 work.","section":"References, Ref. 51"},{"comment":"The caption refers to 'Definition of Production (blue solid line)', which is unclear; it should say the true Cs from the production-based definition.","section":"Figure 12 caption"},{"comment":"The abstract states an improvement of 'about 17%', but comparing the best pointwise F1 of 0.60 with the best 5x5 F1 of 0.73 gives a relative increase of about 22%; please specify the exact baseline and configuration used for the 17% figure.","section":"§IV A 3 and Table VI"},{"comment":"The text refers to 'Appendix V' when discussing classifier comparisons, but the appendix is unnumbered and does not carry a 'V' label; please correct the cross-reference.","section":"§III C 2 and Appendix"}],"recommendation":"major_revision","confidential_remarks":"The core idea is publishable if the evaluation protocol is tightened. The leakage and ensemble-selection concerns are the main barriers; both are fixable with a proper spatial split and a nested selection procedure. I would also ask the authors to be explicit that the reported MSE and F1 numbers concern the projected Smagorinsky coefficient, not the full SGS stress tensor, and to frame the DSM comparison accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Key thing to know: this is a well-organized a priori ML closure study, and its central numerical claims are likely optimistic because the evaluation protocol does not rule out data leakage. The classifier/regressor pipeline for a signed Smagorinsky coefficient is a reasonable practical idea, and the SVD-based geometric invariance is a sensible way to inject local structure. But the reported F1 gains and the very low ensemble MSE should be read as interpolation numbers, not generalization numbers.\n\nWhat is new: the specific combination - a binary classifier to split strong/weak |Cs|, then a regressor for the signed coefficient, using tensor invariants plus SVD of 3x3/5x5 patches on WRF LES of idealized hurricanes at different SSTs - does not appear in exactly this form in prior work. The a priori comparison against dynamic Smagorinsky is appropriate, and the paper is honest that this is Part I, with no a posteriori test yet.\n\nSoft spots, in order of importance. First, the train/test split is not described beyond saying the dataset was randomized. Because each sample is an overlapping patch, point-level randomization means nearby test points are nearly identical to training points, which inflates F1 and deflates MSE. A spatial or temporal split, or at least a separated validation region, is needed. Second, the ensemble aggregates networks that meet an MSE threshold of 1e-8, and the final test MSE is 2.49e-8. The text does not say which data is used for that threshold; if it is the same test set, the reported number is a selection statistic. This must be clarified. Third, the Smagorinsky projection assumption means Cs only captures the component of SGS stress aligned with the strain rate; that is a legitimate stated limitation, but the comparison with DSM is only on that projection and should be framed accordingly. Minor issues: no error bars, no cross-validation details, the 0.001 classification threshold is arbitrary, and the 17% F1 claim depends on the baseline chosen (0.64 to 0.73 is about 14%).\n\nBottom line: the framework is worth engaging, but the evidence as presented is not reliable. A serious referee should ask for a clean evaluation protocol and ideally an a posteriori sanity check. The paper will interest people building data-driven SGS closures for mesoscale atmospheric models, but they should not quote the quantitative claims until the leakage and selection issues are resolved.","headline":"Worth reading for the framework, not for the numbers as they stand: the a priori claims are likely leakage-inflated, but the classifier/regressor pipeline is a sensible contribution.","tokens_in":21557,"tokens_out":3530,"would_cite":false,"duration_ms":32496,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Machine learning can predict the signed Smagorinsky coefficient in hurricane boundary layers, representing both energy cascade and backscatter.","keywords":["large eddy simulation","sub-grid scale stress modeling","Smagorinsky model","energy backscatter","machine learning turbulence closure","tensor invariants","singular value decomposition","hurricane boundary layer"],"falsifier":"Compute the residual between the actual sub-grid stress tensor and its closest Smagorinsky-form reconstruction in regions flagged as backscatter: if the residual energy is not small compared with the total, the signed scalar coefficient cannot carry the claimed physics. In a live simulation, the claim also fails if the model becomes unstable or fails to reproduce the backscatter statistics of the high-fidelity data in a posteriori tests.","tokens_in":20563,"feed_emoji":"🌀","tokens_out":5869,"duration_ms":50070,"temperature":0.7,"pith_summary":"This paper tries to show that a machine-learning sub-grid-scale stress model, built from turbulence invariants and local singular-value features, can predict the signed Smagorinsky coefficient in mesoscale hurricane boundary layer flows well enough to represent both forward energy cascade and energy backscatter. The model splits the task into a classifier that flags regions of strong TKE transfer and a regression ensemble that outputs the signed coefficient; in a priori tests the ensemble matches the true coefficient substantially better than the dynamic Smagorinsky model. The authors' central claim is that embedding geometric invariance through singular values improves classification F1 by about 17 percent over setups without it, and that a scalar Smagorinsky closure with a learned signed coefficient is a practical carrier for backscatter-aware SGS modeling in storm-scale simulations. If right, this gives weather-model large eddy simulation a stable, interpretable route to backscatter without abandoning the Smagorinsky form.","feed_headline":"Neural nets learn the signed Smagorinsky coefficient for hurricane LES","feed_subtitle":"Geometric invariance lifts F1 by about 17 percent, and ensemble regressors beat the dynamic Smagorinsky model in a priori tests.","key_machinery":"The load-bearing object is the Smagorinsky closure with a signed coefficient $\\nu_t = C_s \\Delta^2 |S^*_{ij}|$, where $C_s$ may be negative to model energy backscatter. The true coefficient is defined from the production of TKE, $C_s = (\\tau_{ij}S_{ij} - \\tau_{kk}S_{ii}/3)/(2\\Delta^2 |S^*_{ij}| S^*_{ij}S_{ij})$, which projects the actual SGS stress onto the strain-rate direction. The learning pipeline embeds physical invariance via tensor invariants such as $\\mathrm{tr}(S^2)$ and $\\mathrm{tr}(R^2)$, and geometric invariance via singular values of local $3\\times3$ or $5\\times5$ neighborhoods, which are invariant under rotation and reflection. This mechanism is what lets a pointwise scalar coefficient carry the model: the classifier decides where the coefficient is large, and the ensemble regressor supplies its signed value.","core_discovery":"The discovery the paper argues for is that the signed Smagorinsky coefficient $C_s$, positive for forward TKE cascade and negative for backscatter, can be learned pointwise from local flow data, provided the input encodes physical invariance through tensor invariants of strain and rotation and geometric invariance through singular values of local neighborhoods. Using high-fidelity hurricane boundary layer LES data filtered to a one-kilometer scale, the authors define a true $C_s$ by projecting SGS production onto the Smagorinsky form, then train classifiers and regressors on that target. The resulting ensemble neural networks track the true signed coefficient better than the dynamic Smagorinsky model while avoiding its extreme fluctuations, and the classifier's F1 rises from about 0.60 without geometric invariance to about 0.72 on a 5 by 5 neighborhood, the roughly 17 percent improvement claimed. The paper presents this as a functional SGS model that admits backscatter while remaining a drop-in Smagorinsky closure.","pith_inferences":["The invariance embedding compresses local geometry without assuming a fixed stencil orientation, so the same trained model may transfer across hurricane cases with different storm-relative orientations; the paper does not test this transfer.","If a posteriori tests confirm stability, the classifier-regressor split suggests a cheap gating strategy for hybrid solvers: use standard eddy viscosity in mild regions and machine-learning backscatter only where the classifier says it is needed.","The residual-stress projection issue implies that a tensor-valued machine-learning correction trained on the full stress, rather than only the projected coefficient, could be compared against this model to quantify how much backscatter physics is lost by the Smagorinsky alignment assumption."],"forward_implications":["Classification with 5 by 5 singular-value neighborhoods reaches an F1 score around 0.72, about 17 percent higher than setups without geometric invariance, meaning the model can locate regions where backscatter matters.","Ensemble regressors achieve test mean squared error as low as $2.49 \\times 10^{-8}$ for the signed coefficient, far below the scatter of the dynamic Smagorinsky model in a priori comparisons.","Because the output is a scalar coefficient in Smagorinsky form, the model can be plugged into existing LES and CFD solvers without changing the stress tensor structure.","A signed coefficient permits both forward cascade and backscatter, while the classification stage restricts regression to strong-transfer regions, which the authors argue helps avoid the numerical instability usually associated with anti-diffusion."],"supporting_citations":[{"why":"Provides the Smagorinsky eddy-viscosity formulation that the paper adopts and re-signs as the carrier of the machine learning model.","marker":"[42]"},{"why":"Introduces the dynamic Smagorinsky model used as the baseline comparison for the learned signed coefficient.","marker":"[43]"},{"why":"Supplies the high-fidelity hurricane boundary layer LES dataset at multiple sea surface temperatures used to define the true coefficient and train the models.","marker":"[44]"},{"why":"Demonstrates that accurate backscatter modeling enables stable a posteriori LES, motivating the signed-coefficient approach.","marker":"[18]"},{"why":"Provides radar-based evidence of organized kinetic energy backscatter in the hurricane boundary layer, the physical phenomenon the model targets.","marker":"[41]"},{"why":"Documents the detailed classifier study and model configurations that the present work extends by adding geometric invariance.","marker":"[54]"},{"why":"Reports prior a priori tests without invariance embedding that failed to estimate energy backscatter, establishing the baseline problem this paper addresses.","marker":"[25]"}],"fun_headline_variants":["Invariance-embedded ML learns signed Smagorinsky for hurricane LES","Geometric invariance boosts SGS classifier F1 by 17% in hurricanes","Ensemble neural nets beat dynamic Smagorinsky in a priori storm tests","Hurricane SGS model: ML with physical and geometric invariance gains","Signed Smagorinsky from local invariants predicts backscatter in storms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single signed number multiplying the strain-rate tensor can represent the sub-grid stress in hurricane boundary layers, so any stress part not aligned with that tensor is ignored by both the machine learning model and the dynamic Smagorinsky comparison.","fun_headline_variants_meta":{"raw":{"variants":["Invariance-embedded ML learns signed Smagorinsky for hurricane LES","Geometric invariance boosts SGS classifier F1 by 17% in hurricanes","Ensemble neural nets beat dynamic Smagorinsky in a priori storm tests","Hurricane SGS model: ML with physical and geometric invariance gains","Signed Smagorinsky from local invariants predicts backscatter in storms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1767,"prompt_tokens":1049,"completion_tokens":718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":620}},"tokens_in":665,"tokens_out":718,"duration_ms":6850,"temperature":1.0,"reasoning_tokens":620,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:48:12.633672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the residual between the actual sub-grid stress tensor and its closest Smagorinsky-form reconstruction in regions flagged as backscatter: if the residual energy is not small compared with the total, the signed scalar coefficient cannot carry the claimed physics. In a live simulation, the claim also fails if the model becomes unstable or fails to reproduce the backscatter statistics of the high-fidelity data in a posteriori tests.","supporting_citations":[{"cited_title":"Zhou , author S","cited_arxiv_id":null,"evidence_quote":"Introduces the dynamic Smagorinsky model used as the baseline comparison for the learned signed coefficient."},{"cited_title":"\\ Chang , author A","cited_arxiv_id":null,"evidence_quote":"Demonstrates that accurate backscatter modeling enables stable a posteriori LES, motivating the signed-coefficient approach."},{"cited_title":"Nabavi \\ and\\ author J","cited_arxiv_id":null,"evidence_quote":"Provides radar-based evidence of organized kinetic energy backscatter in the hurricane boundary layer, the physical phenomenon the model targets."},{"cited_title":"Durbin \\ and\\ author B","cited_arxiv_id":null,"evidence_quote":"Documents the detailed classifier study and model configurations that the present work extends by adding geometric invariance."}],"review_version":1}