{"id":"e4668364-d6f0-410d-a4c2-8ca41a11fe13","arxiv_id":"2502.07610","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep-SHAP analysis of a wall-sensing CNN for turbulent channel flow identifies wall pressure as the dominant input and high-importance regions as streak-like clusters of 20 to 120 wall units.","lead":"This paper uses an explainable-AI method, deep-SHAP, to see which wall measurements matter when a neural network predicts velocity fluctuations inside a turbulent channel. It finds wall pressure is the most influential input and that important wall regions form streak-like clusters about 20 to 120 wall units wide, which sets a resolution target for future non-intrusive sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DeepSHAP's independence assumption is load-bearing for the pressure-dominance and cluster-size claims; a correlation-aware attribution test is needed before accepting the sensor-resolution conclusions.","rationale":"The reader's weakest-assumption analysis correctly identifies the deep-SHAP independence approximation as the main risk to the paper's central claims. I agree with that assessment: all of the headline quantitative results—pressure as the most influential input, the identification of high-importance clusters, their streak-like organization, and the 20–120 wall-unit size range used for sensor-resolution guidance—are downstream of the SHAP importance fields. The paper is transparent about the assumption, but the strong correlation between wall pressure and wall-shear stress, which the authors themselves acknowledge in Section 3, is exactly the regime where independent-feature attributions can misallocate credit. The ablation study in Figure 4 is a useful sanity check, but it is not independent of the SHAP ranking because the removed pixels are selected using the same SHAP values; it shows that the top-ranked pixels matter, not that the ranking is unbiased. A well-posed correlation-aware attribution test would settle the issue directly. If the pressure-dominance result survives such a test, the paper's conclusions become considerably more robust; if it does not, the central claim would need to be revised. Given that the data and code are not yet available and that the independence-bias is unquantified, keeping the original CONDITIONAL verdict is appropriate; no adjustment is needed.","tokens_in":13768,"tokens_out":5407,"duration_ms":60494,"concrete_test":"Run the same trained CNN with a correlation-aware attribution method, e.g., conditional Shapley values estimated by Monte Carlo sampling from a Gaussian copula fitted to the DNS wall data, and redo the full pipeline: input ranking, the Figure 4 ablation, cluster segmentation, and size statistics. If pressure no longer yields the largest SHAP values or the clusters no longer fall in the 20–120 wall-unit range, the independence assumption is the cause. If rankings and sizes remain within a small tolerance (say 10%), the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—wall pressure is the most influential input and high-importance clusters are 20–120 wall units—is derived entirely from deep-SHAP importance rankings. Section 2 states that these SHAP values are calculated 'assuming independent features and a linear model' (see Eq. 2.3: phi_i ≈ m_{x_i,f_j}(x_i - E[x_i])). Section 3 later acknowledges a 'strong correlation between the pressure and shear stress fluctuations.' Under input correlation, the independent-feature Shapley value is not the true conditional attribution: the joint contribution of a correlated pressure–shear pattern can be assigned to pressure alone, inflating its apparent influence. Because the same SHAP field is thresholded, clustered, and used to estimate the 80×40 wall-unit (range 20–120) structures, a biased ranking would propagate directly into the sensor-resolution recommendation. The Figure 4 ablation is not a fully independent guard: it removes pixels selected by the SHAP ranking, so it verifies that the selected pixels matter, not that the ranking itself is unbiased. This is the load-bearing soft spot; if correlation-aware attribution reverses the pressure first-place ranking or changes the cluster sizes, the paper's main conclusions do not stand.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the non-intrusive sensing CNN of Guastoni et al. (2021) for a turbulent open channel at Re_tau=180 by applying deep-SHAP to attribute the prediction of velocity fluctuations at y+=15 to the wall-pressure and wall-shear-stress input fields. The central findings are that wall pressure is the most influential input, that high-importance input regions are not the most intense pressure or shear regions but rather have values one to two standard deviations from the mean, that these regions form streak-like clusters with a typical size of about 80 by 40 wall units and a reported range of 20 to 120 wall units, and that this scale sets resolution requirements for non-intrusive sensor arrays. The analysis includes joint histograms of input values and SHAP scores, perturbation experiments that remove top-ranked pixels, and a decile-based clustering of the importance field. The paper also translates the cluster sizes into sensor-resolution constraints for a commercial-aircraft context.","tokens_in":13917,"tokens_out":3628,"duration_ms":36511,"significance":"If the central claims hold, the paper provides practically useful guidance for non-intrusive sensing in wall-bounded turbulence: prioritize pressure-sensor accuracy and resolve structures on the 20-120 wall-unit scale. The framework is coherent, uses an established estimator, and the perturbation experiments in Figure 4 are internally consistent with the SHAP ranking. There is no equation-level circularity: the SHAP values are attributions of a pretrained model, and the ablations are consistency checks rather than fitted constants. The physical interpretation connecting high-importance regions to streaks and sweeps is plausible and consistent with prior work on near-wall structure. However, the main conclusions rest on the deep-SHAP independence assumption applied to inputs that the paper itself states are strongly correlated. Because the same SHAP ranking is thresholded, clustered, and converted into a sensor-resolution recommendation, the independence assumption is load-bearing. The paper would be strengthened by a correlation-aware attribution test and by sensitivity analysis of the clustering thresholds.","major_comments":[{"comment":"The claim that wall pressure is the most influential input is derived from deep-SHAP values computed under the independence assumption stated in Eq. (2.3), yet Section 3 acknowledges a strong correlation between pressure and shear-stress fluctuations. Under input correlation, the independent-feature Shapley value can assign the joint contribution of a correlated pressure-shear pattern to a single input, inflating the apparent influence of pressure. Indeed, the paper explains the pressure dominance as resulting from 'higher errors when the correlation between the input variables is broken,' which is exactly the effect introduced by the independence approximation. Please add a correlation-aware attribution test, for example conditional SHAP with a joint baseline, grouped-coalition Shapley values treating the three wall fields as one coalition, or a permutation test that holds the correlated partner fixed, and report whether pressure remains first in the ranking and whether the cluster-size estimates are unchanged.","section":"Section 2, Eq. (2.3); Section 3"},{"comment":"The perturbation experiment in Figure 4 removes pixels selected by the SHAP ranking, so it verifies that the selected pixels matter for the prediction, but it does not independently validate that the ranking itself is unbiased. If the independence approximation misallocates importance to pressure-dominated pixels, the same misallocation guides the pixel removal and produces the observed MSE increase. Please compare the SHAP-based removal curve against occlusion baselines that remove random pixels, least-important pixels, and pixels selected by an alternative attribution method (e.g., gradient-based or occlusion-based importance), and show that the ranking-specific behavior is not an artifact of correlated input patterns.","section":"Section 3, Figure 4"},{"comment":"The cluster-size statement, including the typical 80 x 40 wall-unit size and the 20-120 wall-unit range, depends on several user-chosen thresholds: the 99th-percentile SHAP threshold, the decile segmentation, and the minimum-area filter S+ = 30^2. The manuscript does not report how the reported size range changes when these thresholds are varied. Since the size range is directly converted into a sensor-resolution recommendation for commercial aircraft, please add a sensitivity analysis over these thresholds and report the resulting spread in cluster sizes, or justify the threshold choices physically.","section":"Section 3, Figure 7 and Conclusions"}],"minor_comments":[{"comment":"The Data availability statement contains the placeholder '[doi]' rather than an actual DOI; please provide the complete link before publication.","section":"Data availability"},{"comment":"The text refers to 'black contours' in the top-right image while the caption refers to 'black dots'; please make the terminology consistent.","section":"Section 3, Figure 6"},{"comment":"The word 'simetrical' should be 'symmetrical'.","section":"Section 3 (near Figure 8)"},{"comment":"The heading 'Aknowledgements' should be 'Acknowledgements'.","section":"Acknowledgements heading"},{"comment":"The phrase 'instantaneous standarized wall pressure' should read 'instantaneous standardized wall pressure'.","section":"Section 3, Eq. (3.2)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of JFM and the central idea is interesting, but the main conclusion depends on the deep-SHAP independence assumption, which the authors themselves note is violated by the correlated inputs. I recommend major revision rather than rejection because the requested tests are feasible within the current framework and could either confirm the pressure-dominance and cluster-size claims or quantify their uncertainty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean, well-presented application of deep-SHAP to a wall-sensing CNN, and it produces a genuinely new result: for the Guastoni et al. model, wall pressure is ranked as the most influential input, and the high-importance regions form streak-like clusters of roughly 80×40 wall units (range 20–120). The segmentation and cluster-size analysis is careful, and the perturbation tests in Figure 4 provide a useful internal consistency check. If the ranking holds, this gives a concrete criterion for sensor placement and resolution, which matters for practical non-intrusive sensing. Credit where due: the methodology is standard, but the application and the physical-size estimates are new, and the paper is clearly written.\n\nThe soft spot is exactly where the stress-test note points. The deep-SHAP values are computed under an independence assumption (Eq. 2.3), and the paper itself admits strong correlation between wall pressure and shear stress. Under correlated inputs, the independent-feature Shapley value can assign joint contributions to one variable, inflating pressure's apparent importance. That is load-bearing because the cluster sizes and the sensor-resolution recommendation are built on the SHAP ranking. The Figure 4 ablation removes pixels chosen by that ranking, so it shows the selected pixels matter, but it does not confirm the ranking is unbiased. The paper would be materially stronger with a correlation-aware attribution test (e.g., conditional SHAP, or integrated gradients with a correlated baseline) to see if pressure still dominates.\n\nOther concerns are more minor. The data and code are referenced with a placeholder DOI, which makes the work hard to verify. The single Reynolds number (Re_tau=180) and the inner-scaling extrapolation to commercial aircraft are speculative; that part should be presented as hypothesis, not conclusion. The free parameters in the clustering (area threshold, importance decile, 99th percentile) are reasonable but should be sensitivity-tested.\n\nOverall, the central architecture of the study is sound and the conclusions are plausible, but the quantitative claims about pressure dominance and cluster sizes need verification before they should be taken as design guidance. The paper deserves a serious referee; I would send it to peer review with a request for the correlation-aware test and the data/code release. I would cite it as an application example, and I might bring it to a reading group focused on interpretable ML in fluids.","headline":"Solid explainable-DL application with a plausible but unverified pressure-dominance ranking; the independence assumption is the load-bearing soft spot.","tokens_in":706,"tokens_out":980,"would_cite":true,"duration_ms":19907,"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":"Wall pressure, not shear stress, is the input a CNN relies on to reconstruct near-wall velocity fluctuations.","keywords":["wall-bounded turbulence","deep learning","explainable deep learning","SHAP values","non-intrusive sensing","wall pressure","wall shear stress","turbulent channel flow"],"falsifier":"A decisive test would be a causal ablation experiment on the same trained CNN: corrupt the wall-pressure field while leaving shear stress intact, and independently corrupt shear stress while leaving pressure intact, then compare the rise in reconstruction error; the paper's claim predicts pressure corruption is markedly more damaging. A second check would recompute the importance ranking with a correlation-aware attribution method and see whether the 20-to-120-wall-unit cluster sizes survive.","tokens_in":13494,"feed_emoji":"🌀","tokens_out":7968,"duration_ms":60328,"temperature":0.7,"pith_summary":"This paper asks which wall measurements a convolutional neural network truly depends on when it reconstructs velocity fluctuations at y+ = 15 inside a turbulent open channel at Re_tau = 180. Using deep-SHAP, an additive feature-attribution method, the authors assign an importance score to every wall grid point of the pressure and shear-stress input fields. They report that wall pressure is the most influential input for all three velocity components, ahead of streamwise and spanwise wall shear stress, and that the high-importance wall regions form streak-like clusters with a typical size of about 80 by 40 wall units and a range of 20 to 120 wall units. If correct, this gives a concrete design rule for non-intrusive sensing: pressure accuracy should be prioritized, and sensor arrays must resolve structures of this scale, which for commercial-aircraft conditions translates to roughly 100 to 600 microns.","feed_headline":"Wall pressure, not shear, dominates deep-learning flow reconstruction","feed_subtitle":"Attribution analysis shows high-impact wall regions are streak-like clusters of 20–120 wall units, setting sensor resolution targets","key_machinery":"The argument is carried by deep-SHAP, a Shapley-value-based attribution method that assigns each input feature a numerical importance for a given prediction. Here the features are the 192 x 192 grid points of wall pressure and streamwise/spanwise wall shear stress, and the output is the predicted velocity fluctuation at y+ = 15; the importance fields are thresholded, segmented into deciles, filtered by a minimum area of S+ = $30^{2}$, and clustered to define coherent high-importance structures. The CNN itself has a receptive field of 15 x 15 grid points, which explains how a small set of important input pixels can influence a larger output region.","core_discovery":"On the paper's own terms, the central discovery is that a CNN trained to estimate near-wall velocity fluctuations from wall quantities is most sensitive to wall pressure, not to wall shear stress, and that the wall regions driving its predictions are moderate-intensity streak-like clusters rather than the strongest fluctuation events. The authors compute deep-SHAP importance fields over 1000 snapshots, rank every input grid point, and show that the top-importance points are concentrated in clusters whose averaged wall pressure lies one to two standard deviations from the mean. These clusters are streamwise-elongated, have a typical area of 80 by 40 wall units and a size range of 20 to 120 wall units, and coincide with low-velocity streaks and ejection-like regions; the shear-stress-driven clusters connect with sweep-like structures. The paper also shows that removing the top 1% of important pressure points distorts predictions over a much larger area, consistent with the network's 15 by 15 receptive field.","pith_inferences":["The paper stops short of testing how reconstruction degrades as sensor spacing exceeds the cluster size; a natural extension would be to downsample the input fields to the inferred cluster scale and measure the error curve.","The independence assumption in the attribution may shift some importance between pressure and shear stress; a correlation-aware attribution or causal ablation would test whether pressure remains dominant.","If the pressure-dominance result carries to higher Reynolds numbers and boundary layers, it would motivate flush-mounted micro-pressure-sensor arrays for closed-loop drag control rather than shear-stress micro-fences.","The coincidence of high-importance wall regions with low-velocity streaks suggests the CNN has implicitly learned the near-wall streak cycle; comparing SHAP-identified regions with an independent streak-detection algorithm would make that link quantitative."],"forward_implications":["Non-intrusive sensing systems for near-wall flow estimation should prioritize wall-pressure accuracy over shear-stress accuracy, since corrupting pressure raises reconstruction error fastest.","A sensor array for this reconstruction task needs to resolve streak-like wall structures of roughly 20 to 120 wall units; in aircraft-like conditions this is about 100 to 600 micrometers, which current wall-stress sensor sizes may not resolve.","The most influential wall regions are not the extreme fluctuation peaks but moderate-intensity regions at one to two standard deviations from the mean, so targeting extreme events alone would miss what the model relies on.","Only the top and bottom importance deciles form clusters large enough to pass the minimum-area filter, meaning the useful information for reconstruction is concentrated in a small fraction of the wall plane.","Because the CNN's receptive field is 15 x 15 grid points, sparse but well-placed pressure sensors can have an outsized effect on the reconstructed velocity field."],"supporting_citations":[{"why":"Provides the CNN architecture and trained model that this paper explains.","marker":"Guastoni et al. 2021"},{"why":"Supplies the deep-SHAP method used to compute per-pixel importance.","marker":"Lundberg & Lee 2017"},{"why":"Defines the Shapley values whose axioms the attribution method satisfies.","marker":"Shapley 1953"},{"why":"Documents the near-wall streaks that the high-importance clusters are identified with.","marker":"Kline et al. 1967"},{"why":"Supplies the minimum-area filter (S+ = 30^2) and the Reynolds-stress structure reference.","marker":"Lozano-Durán et al. 2012"},{"why":"Earlier explainable-deep-learning identification of important regions in wall turbulence, extended here to non-intrusive sensing.","marker":"Cremades et al. 2024"},{"why":"Reports the shear-stress influence on streamwise velocity that the pressure-dominance finding qualifies.","marker":"Cuéllar et al. 2024b"},{"why":"Supplies the shear-sensor resolution figure used to argue current sensors undersample the important clusters.","marker":"Zhe et al. 2005"},{"why":"Provides the 20%-of-structure-size measurement rule used to judge sensor spacing.","marker":"Håkansson et al. 2013"}],"fun_headline_variants":["Wall pressure, not shear, drives deep learning flow sensing","AI explains turbulence sensing: pressure trumps shear","Deep learning pinpoints pressure as key wall sensor signal","Explainable AI: wall pressure leads flow reconstruction","For wall turbulence, pressure beats shear in AI sensing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that deep-SHAP attributions correctly rank input importance even though they are computed assuming independent features and a linear local model, while the actual inputs—wall pressure and wall shear stress—are strongly correlated.","fun_headline_variants_meta":{"raw":{"variants":["Wall pressure, not shear, drives deep learning flow sensing","AI explains turbulence sensing: pressure trumps shear","Deep learning pinpoints pressure as key wall sensor signal","Explainable AI: wall pressure leads flow reconstruction","For wall turbulence, pressure beats shear in AI sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1363,"prompt_tokens":1010,"completion_tokens":353,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":626,"tokens_out":353,"duration_ms":3809,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:09:03.345498+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be a causal ablation experiment on the same trained CNN: corrupt the wall-pressure field while leaving shear stress intact, and independently corrupt shear stress while leaving pressure intact, then compare the rise in reconstruction error; the paper's claim predicts pressure corruption is markedly more damaging. A second check would recompute the importance ranking with a correlation-aware attribution method and see whether the 20-to-120-wall-unit cluster sizes survive.","supporting_citations":[{"cited_title":"Journal of Fluid Mechanics 928 , A27","cited_arxiv_id":null,"evidence_quote":"Provides the CNN architecture and trained model that this paper explains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the deep-SHAP method used to compute per-pixel importance."},{"cited_title":"Contribution to the Theory of Games 2","cited_arxiv_id":null,"evidence_quote":"Defines the Shapley values whose axioms the attribution method satisfies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the near-wall streaks that the high-importance clusters are identified with."},{"cited_title":"Journal of Fluid Mechanics 694 , 100--130","cited_arxiv_id":null,"evidence_quote":"Supplies the minimum-area filter (S+ = 30^2) and the Reynolds-stress structure reference."},{"cited_title":"Nature Communications 15 (1), 3864","cited_arxiv_id":null,"evidence_quote":"Earlier explainable-deep-learning identification of important regions in wall turbulence, extended here to non-intrusive sensing."},{"cited_title":"Journal of Microelectromechanical Systems 14 (1), 167--175","cited_arxiv_id":null,"evidence_quote":"Supplies the shear-sensor resolution figure used to argue current sensors undersample the important clusters."}],"review_version":1}