{"id":"79df3bdc-ea63-4bf6-a7e8-d0bbed39509d","arxiv_id":"2605.26806","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"High-fidelity compressible-flow simulations around a rotating cylinder are used to train supervised machine-learning surrogates that predict aerodynamic loads and the onset of multi-mode instability across a range of Reynolds numbers.","lead":"The paper runs 101 high-fidelity simulations of compressible flow past a rotating cylinder across Reynolds numbers 1000-6000, identifies a flow transition near Re=5650, and trains polynomial, Bayesian, and neural-network models to predict lift, drag, and instability onset. A smart generalist might read it to see how data-driven surrogates can replace repeated expensive fluid simulations for parametric engineering studies.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"Generalization claim for ANN surrogates hinges on unshown density/coverage of the 101 Re samples around the Re=5650 bifurcation","rationale":"The reader's weakest_assumption directly identifies the same sampling-and-generalization risk that the abstract's claims rest on; the full text would need to supply the missing sampling details and hold-out metrics to raise the verdict. No other internal inconsistency is visible from the provided material.","tokens_in":1756,"tokens_out":357,"duration_ms":15372,"concrete_test":"Extract the list of 101 Re values used; perform a leave-one-out or 5-fold cross-validation in which test folds deliberately include at least three Re values within ±200 of 5650; retrain the reported ANN architecture on the complementary folds and report mean relative error on predicted instability onset time and max lift coefficient. If any test-fold error exceeds 8-10%, the generalization claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim requires that the 101 high-fidelity runs form a training distribution dense enough for the ANN (and its hierarchical refinement) to interpolate the transition from periodic shedding to multi-mode states, accurately predict instability onset time, and avoid extrapolation failure at unseen Re. The abstract states a critical bifurcation near Re=5650 and notes that polynomial regression fails near it while ANNs succeed, yet provides no information on Re sampling density, spacing near the transition, or explicit hold-out validation on Re values both below and above the bifurcation. If the 101 points are clustered away from 5650 or lack sufficient post-bifurcation coverage, the reported predictive accuracy for lift and onset time cannot be taken as evidence of reliable surrogate behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper reports high-fidelity simulations of compressible flow past a rotating cylinder for Re = 1000–6000, identifying a bifurcation near Re = 5650 that produces a transition from periodic vortex shedding to multi-mode oscillatory states with nonlinear frequency coupling. From a database of 101 simulations (1 million core hours), it systematically compares polynomial regression (which fails near the bifurcation), Bayesian regression using B-splines and Gaussian RBFs, and artificial neural networks (ANNs). The central claim is that ANNs achieve excellent predictive accuracy for maximum lift coefficient and instability onset time (with reasonable fidelity for drag) and can serve as generative surrogates for unseen Re via a hierarchical refinement strategy.","tokens_in":1957,"tokens_out":657,"duration_ms":28214,"significance":"If the validation and generalization claims are substantiated, the work would offer a useful case study of ML surrogates for compressible flows exhibiting bifurcations and multi-mode coupling, leveraging a substantial high-fidelity database and a systematic method comparison. The hierarchical refinement approach and the contrast between polynomial failure and ANN success near the transition are constructive elements. At present, however, the absence of quantitative error metrics and out-of-sample testing limits the strength of the surrogate reliability conclusion.","major_comments":[{"comment":"Abstract: the assertion of 'excellent predictive accuracy' for maximum lift coefficient and instability onset time is unsupported by any reported quantitative metrics (e.g., MSE, MAE, or R² on a test partition), cross-validation procedure, or explicit test-set performance; the same holds for the 'reasonable fidelity' claim for drag.","section":"Abstract"},{"comment":"Data-driven modeling and results sections: all performance figures are obtained by fitting and evaluating on the identical 101-simulation database; no hold-out set, k-fold cross-validation, or explicit train/test split is described, so statements of prediction 'for unseen Re' remain interpolations within the convex hull of the training points rather than demonstrated generalization.","section":"Data-driven modeling and results sections"},{"comment":"Bifurcation analysis and surrogate evaluation sections: the distribution and local spacing of the 101 Re samples around the critical value Re ≈ 5650 are not reported; without this information it is impossible to verify that the training set is dense enough to capture the transition to multi-mode states or to rule out extrapolation failure in the ANN predictions of onset time and lift.","section":"Bifurcation analysis and surrogate evaluation sections"}],"minor_comments":[{"comment":"The abstract and main text refer to 'polynomial regression' and 'Bayesian regression frameworks' without specifying the exact polynomial degree or the prior choices and hyperparameter settings used for the B-spline and RBF models.","section":"Abstract and §4"},{"comment":"Figure captions and axis labels for the spectral analysis and ANN predictions should explicitly state the number of training points used and whether any points near Re = 5650 were held out.","section":"Figures 8–12"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the thorough and constructive review. The comments correctly identify several areas where quantitative support and validation details are currently missing from the manuscript. We address each major comment below and will revise the paper to incorporate the requested information and analyses.","responses":[{"response":"We agree that the abstract claims lack supporting quantitative metrics and validation details. In the revised manuscript we will report explicit error metrics (MSE, MAE, R²) on a held-out test partition, describe the cross-validation procedure employed, and update the abstract to reflect these quantitative results. We will likewise provide corresponding metrics for the drag coefficient to substantiate the fidelity claim.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the assertion of 'excellent predictive accuracy' for maximum lift coefficient and instability onset time is unsupported by any reported quantitative metrics (e.g., MSE, MAE, or R² on a test partition), cross-validation procedure, or explicit test-set performance; the same holds for the 'reasonable fidelity' claim for drag."},{"response":"The current manuscript indeed reports performance on the full database without an explicit train/test split. We will add k-fold cross-validation results, including metrics on held-out points, and revise the text to clarify that predictions for 'unseen Re' are within the sampled range. Claims of generalization will be adjusted accordingly to reflect the interpolation nature of the tests.","revision_made":"yes","referee_comment":"[Data-driven modeling and results sections] Data-driven modeling and results sections: all performance figures are obtained by fitting and evaluating on the identical 101-simulation database; no hold-out set, k-fold cross-validation, or explicit train/test split is described, so statements of prediction 'for unseen Re' remain interpolations within the convex hull of the training points rather than demonstrated generalization."},{"response":"We will add a figure and accompanying text detailing the distribution and local spacing of the 101 Reynolds-number samples, with emphasis on the region around Re ≈ 5650. This will allow assessment of sampling density near the bifurcation and support evaluation of the ANN predictions in that regime.","revision_made":"yes","referee_comment":"[Bifurcation analysis and surrogate evaluation sections] Bifurcation analysis and surrogate evaluation sections: the distribution and local spacing of the 101 Re samples around the critical value Re ≈ 5650 are not reported; without this information it is impossible to verify that the training set is dense enough to capture the transition to multi-mode states or to rule out extrapolation failure in the ANN predictions of onset time and lift."}],"tokens_in":1550,"tokens_out":564,"duration_ms":32694,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper generates 101 high-fidelity simulations of compressible flow past a rotating cylinder from Re 1000 to 6000, identifies a bifurcation near 5650 where periodic shedding gives way to multi-mode states, and then fits polynomial, Bayesian spline/RBF, and ANN models to the resulting lift, drag, and onset-time data. The ANNs are shown to handle the nonlinear trends near the transition better than the simpler fits.\n\nThe simulation database and the spectral analysis of the load signals are the clearest contributions. They provide a concrete record of how frequency content and amplitude modulation change across the bifurcation, and the side-by-side comparison of the three regression families illustrates a practical difference in this specific flow.\n\nThe soft spots are the missing quantitative details. No error metrics, cross-validation scheme, or description of Re-point spacing appear in the abstract, and all performance statements read as in-sample fits. Without knowing how densely the 101 runs cover the region around 5650, it is difficult to judge whether the ANN advantage reflects genuine generalization or simply better interpolation on well-sampled data. The claim that the networks can reconstruct behavior at unseen Re therefore rests on unshown coverage.\n\nThis work is for people who run similar cylinder-wake studies or who need reference data for surrogate modeling in rotating-body aerodynamics. A reader looking for new algorithms will find little, but someone who wants to see how off-the-shelf methods perform on real compressible CFD output will get usable information.\n\nIt deserves peer review. The underlying simulations represent real effort, the physics is presented directly, and the method comparison is falsifiable once the validation numbers and sampling plot are supplied. A referee can request those items without the paper being fundamentally broken.","headline":"This is a straightforward application of standard regression methods to a new set of 101 compressible rotating-cylinder simulations, with the main value in the generated data and the documented bifurcation rather than any new modeling technique.","tokens_in":2415,"tokens_out":434,"would_cite":false,"duration_ms":37037,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Artificial neural networks trained on 101 high-fidelity simulations serve as reliable surrogates for predicting lift, drag, and instability onset in compressible flow past a rotating cylinder.","keywords":["compressible flow","rotating cylinder","artificial neural networks","surrogate modeling","vortex shedding","flow bifurcation","Reynolds number","aerodynamic loads"],"falsifier":"Large prediction errors on drag coefficient or onset time when the trained network is tested at a Reynolds number between 1000 and 6000 that was deliberately withheld from the original 101 simulations would show the surrogate claim does not hold.","tokens_in":2664,"feed_emoji":"🌀","tokens_out":638,"duration_ms":28771,"temperature":0.7,"pith_summary":"The paper examines compressible flow past a rapidly rotating cylinder across Reynolds numbers from 1000 to 6000 using 101 expensive numerical simulations. It identifies a shift from periodic vortex shedding to multi-mode oscillations with a bifurcation near Re = 5650. Polynomial and Bayesian regression methods are compared as data-driven fits, but artificial neural networks are shown to deliver high accuracy on maximum lift and instability timing while handling the harder drag signal adequately. The networks are also tested as generative models that can fill in behavior at Reynolds numbers absent from the training data.","feed_headline":"Neural nets predict cylinder lift and instability from 101 runs","feed_subtitle":"They generalize across Reynolds numbers and capture the shift to multi-mode oscillations near 5650.","key_machinery":"Artificial neural networks developed as high-capacity surrogate models, trained on the 101-simulation database and evaluated both for regression and for generative reconstruction at unseen Reynolds numbers.","core_discovery":"When trained on data from 101 high-fidelity compressible flow simulations, artificial neural networks achieve excellent predictive accuracy for maximum lift coefficient and instability onset time, while maintaining reasonable fidelity for the more challenging drag coefficient, allowing them to function as efficient and reliable surrogates for the nonlinear fluid problem that includes bifurcation and mode coupling.","pith_inferences":["Similar ANN surrogates could lower the cost of exploring rotating-body flows at additional Mach numbers or cylinder aspect ratios.","The hierarchical refinement step might be adapted to combine data from simulations at different grid resolutions.","Direct comparison against wind-tunnel measurements at the same Reynolds numbers would test whether simulation-trained networks transfer to experiments."],"forward_implications":["ANNs outperform polynomial regression near the bifurcation point where localized fluctuations appear.","Bayesian spline models improve uncertainty estimates but are outpaced by ANNs in overall capacity.","The trained network can reconstruct lift and drag signals at Reynolds numbers not present in the training set.","The same supervised learning approach is positioned for use on other fluid problems with strong nonlinear dependencies."],"fun_headline_variants":["ANNs predict lift and onset in rotating cylinder flow","ML models capture bifurcation at Re 5650 in cylinder flow","Data driven surrogates for compressible rotating cylinder flow","Neural surrogates model nonlinear cylinder wake dynamics"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The 101 high-fidelity simulations form a sufficiently dense and representative set that lets the networks generalize accurately to new Reynolds numbers and capture the bifurcation without overfitting.","fun_headline_variants_meta":{"raw":{"variants":["ANNs predict lift and onset in rotating cylinder flow","ML models capture bifurcation at Re 5650 in cylinder flow","Data driven surrogates for compressible rotating cylinder flow","Neural surrogates model nonlinear cylinder wake dynamics"]},"model":"grok-4.3","cost_usd":0.006471,"raw_usage":{"total_tokens":3036,"prompt_tokens":680,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":64712000,"prompt_tokens_details":{"text_tokens":680,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2296,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":680,"tokens_out":60,"duration_ms":26407,"temperature":1.0,"reasoning_tokens":2296,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T15:53:51.189005+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Large prediction errors on drag coefficient or onset time when the trained network is tested at a Reynolds number between 1000 and 6000 that was deliberately withheld from the original 101 simulations would show the surrogate claim does not hold.","supporting_citations":[],"review_version":1}