{"id":"6dadcab6-847c-4aa4-8aef-38500c407442","arxiv_id":"1908.07307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A GAN trained on 30% of the TPU interference database predicts mean and fluctuating pressure coefficients on a tall building for unseen neighboring-building positions and wind angles.","lead":"This paper trains machine learning models on wind tunnel data to predict wind pressures on a tall building when a neighboring building creates interference. The best model, a generative adversarial network, needs only 30% of the test cases, which could substantially reduce wind tunnel testing costs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 30%-sufficiency threshold rests on a single random split with no error bars; repeated-seed and baseline comparisons are needed before claiming 70% of wind tunnel tests can be saved.","rationale":"The paper has a credible engineering goal, and the GANs architecture is described in sufficient detail to be reproduced; the reported R2 values on one test split suggest the method has promise. However, the load-bearing step in the argument is the extrapolation from one random split to a general 30% rule. The dataset's structured nature makes random-split variance a first-order concern: a 30% sample of 2664 cases may unevenly cover the 37 locations and 72 angles, and the subsequent 20% test split is drawn from the same sample, so there is no independent check of generalization. The 6 final validation cases are too few, and showing only 2 of them weakens the evidence. I also note the missing interpolation baseline because the paper explicitly criticizes linear interpolation for untested scenarios; demonstrating GANs superiority over such a baseline is necessary for the practical claim. These are conditions on the evidence rather than fatal flaws, so the CONDITIONAL verdict remains appropriate. The limitations are omissions of evaluation rigor rather than internal contradictions, and there is no indication of intentional misrepresentation.","tokens_in":16237,"tokens_out":7569,"duration_ms":79986,"concrete_test":"Retrain the 30%-portion GANs with 10 different random seeds (and, separately, with 5-fold stratified splits by interfering-building location) using the same TTV procedure, and report mean±std of test R2 and MSE for mean and fluctuating pressures on the 20% test split and on the 6 held-out cases. Additionally, fit a simple interpolation baseline (e.g., scipy griddata linear or kriging in Sx, Sy, θ) to the same 30% training cases and evaluate it on the same test/held-out cases. If any seed's fluctuating-pressure R2 falls below about 0.85, or if the interpolation baseline matches the GANs within a few percent, the '30% is sufficient / 70% savings' claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim—that a GANs model trained on 30% of the TPU dataset is equivalent to the full 2664-case set—is supported only by one random selection of 6 held-out cases and one random 30% subset (Sections 2 and 4.3). Figures 11 and 12 plot a single MSE/R2 trajectory for each data portion, with no repeated seeds, confidence intervals, or stratified splits. Because the data lie on a structured grid (37 interfering-building locations × 72 wind angles, Section 2), a single favorable draw can make the 20% test split inside the 30% subset look easy and can exaggerate the 'diminishing returns' beyond 30%. The final validation uses only 6 cases, of which only 2 are displayed; this is too small to establish the headline 'any unseen interference conditions.' Additionally, no non-ML baseline (e.g., linear or kriging interpolation in Sx, Sy, θ) is compared, so the claim that the GANs technique specifically enables the 70% savings is not established. If the R2 on a different split drops, the recommended 30% threshold and the 70% savings statement would have to be revised.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes using machine learning to predict mean and fluctuating wind pressure coefficients on a tall principal building under interference from an identical neighboring building, using the Tokyo Polytechnic University aerodynamic database (2664 cases; 37 interferer locations × 72 wind angles). Four models—decision tree regressor, random forest, XGBoost, and a custom two-stream GAN with a global-local generator and patch discriminator—are compared on a 30% subset of the data; the GAN achieves the highest R² on the inner test set. The authors then train GANs on 10–90% of the data and, based on MSE/R² curves and two visual comparisons on six held-out cases, conclude that 30% suffices for accurate prediction and that 70% of wind tunnel tests could be saved.","tokens_in":16492,"tokens_out":6795,"duration_ms":64101,"significance":"The practical goal is attractive: if a generative model can learn pressure distributions from a limited set of wind-tunnel tests, the cost of interference studies could be reduced substantially. The paper is among the early applications of GANs to wind-pressure prediction, uses a public benchmark database, and includes careful hyperparameter optimization for classical baselines; the four-model comparison and visual pressure-map outputs are useful. However, the headline quantitative claim—that 30% of the data is equivalent to the full set and that 70% of tests can be saved—rests on a single random split, a small hold-out set, and no comparison to classical interpolation. The significance is therefore conditional on additional validation experiments.","major_comments":[{"comment":"The 30% threshold and the accompanying \"70% savings\" claim are based on a single random split with no error bars or repeated-seed trials. The MSE/R² trajectories show one run per data portion; because the test set for each portion is the 20% held out from that same portion, the curves conflate training-set size with changes in test-set composition and size. To support the central claim, the authors should fix an independent test set, repeat the random selection over multiple seeds, and report mean ± standard deviation (or confidence intervals) for each portion.","section":"§4.3, Figs. 11-12"},{"comment":"The final validation uses only six held-out cases, and only two are displayed \"due to page limit\"; the manuscript asserts rather than demonstrates that these cases span the relevant interference conditions. No per-case quantitative errors are reported for the six cases, and no coverage analysis (e.g., spanning critical locations Sx, Sy and wind angles) is provided. This is too thin to support the claim of accurate prediction under \"any unseen interference conditions.\"","section":"§4.1 and §4.3"},{"comment":"The claim that GANs specifically enable the 70% savings is not tested against a non-ML baseline. The introduction dismisses linear interpolation as \"questionable,\" but no quantitative comparison with linear/kriging interpolation or other regressors is given. Without such a baseline, the observed accuracy at 30% data may reflect the smoothness of the pressure field rather than a property of the GAN architecture.","section":"§5 and Introduction"},{"comment":"The equivalence claim—that the GANs model based on 30% of the dataset can be considered equivalent to the entire 2664 dataset—is not directly tested. The authors never train a GAN on the full dataset and compare it on the same fixed hold-out set; the 10–90% curves use changing inner test sets and only two visual comparisons between 30% and 50% models. A direct comparison of GAN-30% and GAN-100% on a fixed test set, with a statistical equivalence or non-inferiority test, is needed.","section":"§4.3 and Conclusions"},{"comment":"All evaluated inputs are discrete grid points from the 37 locations × 72 angles; no off-grid condition (e.g., Sx = 5.5B or θ = 57.5°) is tested, although the stated goal includes \"untested interference conditions\" and \"high-resolution interference factor maps.\" The interpolation capability of the GAN in continuous input space therefore remains unvalidated.","section":"§4.3 and Introduction"}],"minor_comments":[{"comment":"\"Principle building\" appears in Section 2 and the Conclusions; it should be \"principal building.\"","section":"Section 2"},{"comment":"\"Building breaths\" should be \"building breadths.\"","section":"Introduction"},{"comment":"Mt and Ft are not defined before use; the authors should define the ground-truth mean and fluctuating pressure maps and correct the garbled summation notation \"5É\" in the displayed equation.","section":"§3.4.3, Eq. (2)"},{"comment":"The opening sentence refers to \"2 unseen interference conditions\" for the validation; this should read \"6 held-out cases (two shown).\"","section":"§4.2"},{"comment":"The statement that the 30%-based GANs model is \"superior to the full dataset\" is conceptually misstated; the intended claim is presumably that it achieves accuracy comparable to a model trained on the full dataset, which is not yet demonstrated.","section":"Conclusions"},{"comment":"The abstract and discussion say \"30% of the entire 2664 dataset,\" whereas Section 2 says 30% of the remaining 2558 cases after removing the six hold-out cases; this inconsistency should be corrected.","section":"Abstract/Section 5"}],"recommendation":"major_revision","confidential_remarks":"I see no circularity or fabrication concerns; the paper uses a public database and standard supervised learning. The main risk is overclaiming from a single split; if the authors cannot provide repeated-split or full-data comparisons, the 70% savings statement should be substantially softened. Also, the paper would benefit from releasing the exact train/validation/test split to make the results reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper applies GANs to a genuine wind engineering problem: predicting mean and fluctuating pressure maps on a tall building under interference from a neighboring building, using the public TPU database. The two-stream global-local generator with patch discriminator is a sensible design for this task, and the comparison against decision trees, random forests, and XGBoost is fair. The GAN clearly wins, and it is plausible why: it treats pressure maps as images and learns spatial structure, whereas the other methods predict pointwise. The data-portion study (10% to 90%) is a useful practical question, and the 30% elbow is suggestive.\n\nThat said, the headline claim—that 30% of the data suffices to replace the full set and save 70% of wind tunnel tests—is not supported by the evidence presented. The entire quantitative case rests on a single random split: one 30% subset, one 20% test split within it, and six held-out cases (two shown). No repeated seeds, no confidence intervals, no stratified sampling. With only 37 locations and 72 angles on a structured grid, a favorable draw could easily produce an optimistic elbow. The paper needs repeated-seed experiments and, ideally, a few different 30% draws, to show that the threshold is stable.\n\nMore importantly, there is no baseline against simple interpolation in the parameter space (e.g., linear or kriging in Sx, Sy, and wind angle). The paper dismisses linear interpolation as inadequate, but it never tests it. If cheap interpolation matches the GAN at 30% data, the \"savings\" claim collapses. That comparison is essential before making the 70% claim.\n\nThe six validation cases are too few to support 'any unseen interference conditions.' They are all within the training grid, so this is interpolation, not extrapolation. The conclusion that the 30% model is 'equivalent to the entire 2664 dataset' overstates what a single test split shows.\n\nNone of this is fatal to the core idea. The architecture is sound, the dataset is public, and the methodology is mostly transparent. This is a credible engineering application that deserves peer review—but with major revisions required: repeated splits, error bars, an interpolation baseline, and more careful claims about what 'unseen' means. I would bring it to a reading group to discuss how ML evaluation practices can mislead in applied domains, but I would not cite the 30% claim as established.\n\nRecommended verdict: major revision.","headline":"Solid GAN application to wind pressure prediction, but the 30%-sufficiency/70%-savings claim needs repeated splits and an interpolation baseline before it can be trusted.","tokens_in":16999,"tokens_out":3483,"would_cite":true,"duration_ms":35127,"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":"A GAN trained on only 30% of a 2,664-case wind-tunnel pressure database predicts untested building-interference configurations as accurately as the full dataset.","keywords":["wind pressure coefficients","interference effects","tall buildings","generative adversarial networks","machine learning","wind tunnel testing","pressure map prediction"],"falsifier":"Take the same 30% training budget but hold out a structured slice of the configuration space—for example, all cases with the interfering building in a particular upstream region or all wind angles in a 30-degree band—and then evaluate the trained GAN on that slice. If the coefficient of determination drops substantially below the reported 0.988 and 0.924, or the pressure maps miss the known separation regions, the claim of accurate prediction under any unseen interference condition would be refuted for that part of the space.","tokens_in":16056,"feed_emoji":"🌬️","tokens_out":9485,"duration_ms":89064,"temperature":0.7,"pith_summary":"The paper attacks a practical conflict: understanding how a neighboring tall building changes wind pressures on a principal building normally requires hundreds of wind-tunnel runs, but budgets force engineers to test only a few layouts. The authors claim that a generative adversarial network can close this gap. Trained on only 30% of the measured pressure cases in a public database, their GAN reproduces both mean and fluctuating pressure coefficients on the building faces for unseen interferer positions and wind angles, with coefficients of determination of 0.988 for mean and 0.924 for fluctuating pressures. The key practical claim is that those 30% of cases are enough: beyond 30%, more data adds almost nothing, so 70% of wind-tunnel testing could be saved. If true, the result turns a data-hungry experimental program into a one-time training set plus a fast learned surrogate.","feed_headline":"GAN predicts building pressures from 30% of wind-tunnel tests","feed_subtitle":"One model trained on a third of the cases matches full wind-tunnel accuracy, saving 70% of tests.","key_machinery":"The load-bearing object is a conditional GAN with a global-local generator and a patch discriminator. The condition is the triplet (Sx,Sy,θ)—the interfering building's position and the wind angle—and the output is a paired set of pressure maps, mean and fluctuating, on the building's four faces. Fully connected layers produce coarse initial maps, then five residual blocks with side outputs refine local detail, while two patch discriminators classify small 3×4 regions as real or fake; this adversarial loss enforces high-frequency spatial structure. The key property is that the network learns pressure fields as images, so spatial correlations across faces and around separation regions are part of the prediction, which is why it outperforms pointwise tree-based and boosting models.","core_discovery":"The central discovery is that a purpose-built conditional GAN, rather than a pointwise regressor, can serve as an emulator of an entire wind-pressure database. Given the interferer's coordinates relative to the principal building and the wind attack angle, the generator outputs two 9×28 maps—one for mean pressure coefficients and one for fluctuating pressure coefficients—unwrapped across the four faces. The adversarial discriminator, applied to local patches, preserves the sharp spatial structure of separation bubbles and reattachment zones that average losses would blur. In head-to-head tests the GAN beats decision tree, random forest, and XGBoost, and it maintains its accuracy when trained on only 30% of the database; the paper therefore states that the 30%-trained model is equivalent to the full 2,664-case dataset. The practical conclusion drawn is that 70% of wind-tunnel test cases can be omitted.","pith_inferences":["Extrapolating beyond the demonstrated geometry, the 30% threshold should not be assumed to hold for unequal building heights, different cross-sections, or multiple interfering buildings; each new configuration class would need its own training set and saturation curve.","A structured hold-out test—reserving a whole spatial region or wind-angle band rather than a random 20%—would be a stricter check of whether the GAN is learning the physics of interference or memorizing neighborhoods in the training grid.","The same generator could be placed inside an optimization loop to find worst-case interferer locations for cladding or crosswind response, converting a predictive model into a design-search tool.","The near-flat accuracy gain beyond 30% hints that the pressure field varies smoothly with interferer position and wind angle in this two-building configuration, so the model's usefulness depends on interpolation within the tested envelope rather than extrapolation beyond it."],"forward_implications":["Wind engineers could test roughly one in three interference configurations and generate the remaining two thirds computationally, cutting campaign time and cost.","The model can be re-evaluated at any continuous interferer position and wind angle, so interference factor maps for forces and moments can be produced at higher resolution than the discrete 5-degree and discrete-position grid of the original tests.","The same image-to-image training recipe is a template for predicting other spatially distributed wind effects, such as cladding pressures or roof pressures, from limited experimental data.","The data-volume curve, with accuracy saturating near 30%, gives a practical rule for sizing future wind-tunnel matrices before a campaign begins."],"supporting_citations":[{"why":"Supplies the adversarial training objective that the generator-discriminator pair is built on.","marker":"Goodfellow et al. (2014)"},{"why":"Documents the interference-pressure testing program that provides the database used for training and validation.","marker":"Kim et al. (2011)"},{"why":"Extends the same test program to aerodynamic wind forces and is cited as part of the database documentation.","marker":"Kim et al. (2015)"},{"why":"Gives the rationale for using a learned discriminator as a texture or style loss to avoid blurry outputs from L1 and L2 training.","marker":"Larsen et al. (2016)"},{"why":"Provides the residual-block architecture used in the generator's local refinement stream.","marker":"He et al. (2016)"},{"why":"Provides the side-output mechanism used to supervise intermediate pressure maps in the generator.","marker":"Xie & Tu (2015)"},{"why":"Supplies the training-testing-validation evaluation protocol that structures the experiments.","marker":"Reich & Barai (1999)"},{"why":"Supplies the random-forest baseline that the GAN must outperform in the model comparison.","marker":"Breiman (2001)"},{"why":"Supplies the XGBoost baseline that the GAN must outperform in the model comparison.","marker":"Chen & Guestrin (2016)"}],"fun_headline_variants":["GAN maps wind pressures from 30% of wind-tunnel data","70% of wind-tunnel tests cut by GAN pressure predictor","Generative model learns building pressures from 30% cases","Tall-building wind loads predicted from third of tests","AI emulator saves 70% of wind-tunnel experiments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the six reserved validation cases and the 20% test split inside one randomly chosen 30% training subset fairly represent all 2,664 possible interferer positions and wind angles, so the high agreement scores mean the model can interpolate anywhere in the configuration space, not just near the cases it was shown.","fun_headline_variants_meta":{"raw":{"variants":["GAN maps wind pressures from 30% of wind-tunnel data","70% of wind-tunnel tests cut by GAN pressure predictor","Generative model learns building pressures from 30% cases","Tall-building wind loads predicted from third of tests","AI emulator saves 70% of wind-tunnel experiments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2643,"prompt_tokens":957,"completion_tokens":1686,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1602}},"tokens_in":573,"tokens_out":1686,"duration_ms":10743,"temperature":1.0,"reasoning_tokens":1602,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:20:15.803870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same 30% training budget but hold out a structured slice of the configuration space—for example, all cases with the interfering building in a particular upstream region or all wind angles in a 30-degree band—and then evaluate the trained GAN on that slice. If the coefficient of determination drops substantially below the reported 0.988 and 0.924, or the pressure maps miss the known separation regions, the claim of accurate prediction under any unseen interference condition would be refuted for that part of the space.","supporting_citations":[{"cited_title":", Pouget-Abadie, J","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial training objective that the generator-discriminator pair is built on."},{"cited_title":", Tamura, Y","cited_arxiv_id":null,"evidence_quote":"Documents the interference-pressure testing program that provides the database used for training and validation."},{"cited_title":", Tamura, Y","cited_arxiv_id":null,"evidence_quote":"Extends the same test program to aerodynamic wind forces and is cited as part of the database documentation."},{"cited_title":", S nderby, S K","cited_arxiv_id":null,"evidence_quote":"Gives the rationale for using a learned discriminator as a texture or style loss to avoid blurry outputs from L1 and L2 training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the side-output mechanism used to supervise intermediate pressure maps in the generator."},{"cited_title":"\\ Barai, S","cited_arxiv_id":null,"evidence_quote":"Supplies the training-testing-validation evaluation protocol that structures the experiments."},{"cited_title":"APACrefauthors \\ 2001","cited_arxiv_id":null,"evidence_quote":"Supplies the random-forest baseline that the GAN must outperform in the model comparison."},{"cited_title":"\\ Guestrin, C","cited_arxiv_id":null,"evidence_quote":"Supplies the XGBoost baseline that the GAN must outperform in the model comparison."}],"review_version":1}