{"id":"fe199b2a-71c9-436c-a9ea-7e962dcfb2fd","arxiv_id":"2508.18703","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A data-driven quasi-steady model, including advance-ratio, spanwise-velocity and rotational-Wagner terms, predicts CFD forces of flapping wings in forward and maneuvering flight better than the conventional model.","lead":"Researchers used a data-driven search over 5,500 candidate formulas to refine the quasi-steady model of flapping-wing aerodynamics. The upgraded model cuts force-prediction error on separate test flights from about 29% to 11-16%, and it adds three mechanisms missing from the old model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training data set body angular velocity to zero, but the only maneuvering test (fruit-fly turn) has strong body rotation; the maneuvering generalization may be an untested extrapolation.","rationale":"The reader's weakest assumption is exactly the load-bearing concern: the training data set body angular velocity to zero, while the fruit-fly turning validation contains substantial body rotation. This matters because the central claim includes maneuvering flight, and the model equations do incorporate omega_b at test time. The paper's own residual analysis focuses on vortex breakdown and does not address whether the training distribution gap contributes to the turn-segment error. The concern is not fatal: the improvement in hovering and forward flight is supported, and the random-model control provides some evidence against pure overfitting, but the maneuvering leg of the headline claim rests on an untested extrapolation. A retraining or augmentation experiment would settle it. Since the reader already conditioned on addressable limitations, my read does not move the verdict: it remains CONDITIONAL.","tokens_in":33708,"tokens_out":5694,"duration_ms":71453,"concrete_test":"Re-run the discovery and coefficient-fitting pipeline on a training set augmented with nonzero body angular velocity (e.g., sinusoidal roll/pitch/yaw waveforms spanning the measured fruit-fly turn rates), then evaluate on the original fruit-fly maneuver. If the turn-segment RMSE drops materially or body-rotation-related functions enter the selected set, the zero-omega_b assumption is violated and the current maneuvering error is partly an extrapolation artifact; if the error is unchanged, the assumption holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim explicitly includes maneuvering flight, yet the only maneuvering validation is the fruit-fly evasive sequence in Fig. 4(c,d), which contains large body roll/pitch/yaw. In Sec. II B 1 the training data were generated with omega_b = 0, under the stated assumption that variations in theta_pos, theta_ele, and theta_fea sufficiently represent the influence of body rotation. At test time, however, the model uses Eq. 4 (omega_w = omega_b + omega_b,w) and Eq. 6 (v_w = v_b + omega_w x p), so velocity and angular-velocity patterns induced by body rotation enter the kinematic features. No training sample contained those patterns, so the coefficients and selected mechanisms were optimized in a regime disjoint from the maneuver being claimed. The assumption that wing-angle variation can substitute for body rotation is plausible for orientation, but not obviously for the velocity field: body rotation adds a p x omega_b term whose spatial structure differs from wing-relative rotation. If this assumption fails, the low reported error on the fruit-fly turn is partly extrapolation, and the 'maneuvering flight' leg of the headline claim is overstated. Figure 9 attributes large residuals to vortex breakdown, but it does not test whether the zero-omega_b training gap also contributes. No ablation or body-rotation-only training set is provided, so the concern remains unresolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a data-driven extension of the quasi-steady model (QSM) for flapping-wing aerodynamics. The authors build a large synthetic training set of 1,200 wingbeats with CFD-computed forces, construct a library of 5,548 candidate kinematic functions, and use a RIDGE/recursive-feature-elimination procedure to identify 100 robust candidate terms. From these, three mechanisms are manually interpreted and formulated: the advance-ratio effect, the spanwise kinematic-velocity effect, and the rotational Wagner effect. Incorporating these into a 34-term QSM, they report substantially reduced prediction errors against CFD for measured hawkmoth (hovering and forward flight) and fruit-fly (evasive maneuvering) kinematics, with RMSE values of 11.0–16.0% (hawkmoth) and 6.7–15.4% (fruit fly), compared with about 29% for the conventional QSM. The authors also provide grid-independence, robot validation, and random-model comparison as supporting evidence.","tokens_in":34139,"tokens_out":5204,"duration_ms":68617,"significance":"If the result holds, the paper makes a useful contribution: it provides explicit algebraic formulations for three mechanisms that were previously recognized only qualitatively, and it demonstrates out-of-sample prediction on measured kinematics that were not used in training. The random-model comparison in Fig. 7(c) is a strong addition, as it shows that the selected functions generalize better across species and flight modes than arbitrary low-order kinematic terms. The validation against robot experiments and the grid-independence study also increase confidence in the CFD ground truth. The main qualification is that the 'maneuvering flight' leg of the headline claim is validated by a single fruit-fly sequence that contains strong body rotation, while the training data deliberately set body angular velocity to zero. This gap does not undermine the forward-flight or hovering results, but it leaves the maneuvering generalization less secure than the abstract implies.","major_comments":[{"comment":"The training data set body angular velocity to zero (Sec. II B 1), under the assumption that variations in θpos, θele, and θfea sufficiently represent the influence of body rotation. The only maneuvering test, however, is the fruit-fly evasive sequence, which contains large roll, pitch, and yaw (Fig. 4d). At test time the model evaluates Eq. (4) and Eq. (6) with nonzero ωb, so the wing velocity and angular-velocity features that enter the kinematic library are induced by body rotation. No training sample contained such patterns. The spatial structure of the ωb×p term differs from the wing-angle variations used in training, so the assumption is stronger than a mere change of basis. The low reported error on the fruit-fly turn may therefore include an extrapolation component, and the claim of accurate prediction 'across maneuvering flight' is not fully established. I recommend either addin","section":"§II B 1, Eqs. (4) and (6), Fig. 4(d)"},{"comment":"The headline error ranges (11.0–16.0% for hawkmoth, 6.7–15.4% for fruit fly) are point estimates over a small number of measured sequences. The random-model comparison in Fig. 7(c) is qualitatively convincing, but no confidence intervals or repeated train/test splits are reported, so the reader cannot assess whether the improvement over the conventional QSM is statistically significant for each flight mode. This is particularly relevant for the fruit-fly maneuver, where only one evasive sequence is used. Additional independent maneuvers or a leave-one-sequence-out analysis would materially strengthen the claim that the model generalizes across maneuvering flight.","section":"§III C, Fig. 7(c)"}],"minor_comments":[{"comment":"Only eight of the 100 selected functions are shown; the text refers to the supplementary material for the full list. Please include the complete table in the main text or ensure the supplementary URL is available and clearly referenced, as the reproducibility of the feature-selection step depends on it.","section":"Table II"},{"comment":"In the definition of ftc,sp,Wg, the final factor is cos(αx), whereas the corresponding spanwise-chord expressions use cos(αy). Please check whether this is a typo; if intentional, clarify why the chordwise angle of attack appears in the spanwise Wagner term.","section":"Appendix D1, Eq. (D1)"},{"comment":"The derivation of the Wagner correction uses a mean-value argument (introduction of teq) without stating the required smoothness/regularity conditions on Γqs(t). State the assumptions or provide a numerical verification that the approximation is valid over the range of accelerations encountered.","section":"Appendix D3, Eqs. (D10)–(D13)"},{"comment":"The heading reads 'Coeffecients' and should be 'Coefficients'.","section":"Appendix D2, heading"},{"comment":"The horizontal axis in Fig. 9(b) is not labeled in the caption; please specify that the error is averaged over wingbeat phase or over successive stroke cycles, as appropriate.","section":"Fig. 9(b)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically solid in its core methodology and the forward-flight/hovering validation is convincing. The main concern is the body-rotation gap in the training data, which directly affects the maneuvering claim in the abstract and conclusion. This is fixable either by adding body-rotation training data or by softening the claim, so I do not recommend rejection. The paper seems well within scope for physics of fluids / bio-inspired flight modeling."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThis paper is worth your time if you work on quasi-steady models of flapping wings. It formalizes three mechanisms—advance-ratio decomposition, spanwise kinematic velocity, and a rotational Wagner effect—and integrates them into a QSM that cuts prediction error roughly in half on measured hawkmoth and fruit fly kinematics. The out-of-sample validation is real: the measured data were not used in training, and the CFD ground truth is checked against robot experiments and grid independence. The random-model control is a nice touch, showing the selected functions generalize better than arbitrary complexity.\n\nThe soft spot is the maneuvering claim. The training data set body angular velocity to zero (Sec. II B 1), so all learned coefficients see only wing-relative rotation. The one maneuvering test, the fruit-fly evasive sequence, has strong body roll/pitch/yaw. At test time the model uses full ω_w = ω_b + ω_b,w and v_w = v_b + ω_w × p, so the body rotation introduces velocity patterns (ω_b × p) never present in training. The low error on that turn may therefore be partly extrapolation, not validated generalization. The authors attribute residuals to vortex breakdown (Fig. 9), but they do not test the training gap—no ablation, no body-rotation-only training set. I don't think this invalidates the central result for hover and forward flight, where the domain gap is absent, but it does mean the 'maneuvering flight' leg is overstated.\n\nOther concerns are minor: many coefficients are fit to the same data used for selection (the random-model control mitigates but cannot fully exclude overfitting), there is no code or data release, and the Wagner teq is set to an eighth of a stroke based on one reference, which is a bit ad hoc. All addressable.\n\nOverall: a solid, useful contribution for the QSM community. It deserves a serious referee. I'd recommend engaging—but ask for an ablation or training set with body rotation to support the maneuvering claim, and release code/data.","headline":"Useful data-driven QSM refinement with strong out-of-sample gains for hover and forward flight, but the maneuvering validation rests on a single test where training excluded body rotation, so that part of the claim is extrapolation.","tokens_in":34556,"tokens_out":3746,"would_cite":true,"duration_ms":41264,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Three added mechanisms cut flapping-wing force error from about 29% to 7–16%.","keywords":["flapping-wing aerodynamics","quasi-steady model","data-driven model discovery","advance ratio","spanwise flow","rotational Wagner effect","insect flight","aerodynamic force prediction"],"falsifier":"Train the same sparse-regression pipeline on data that include nonzero body roll, pitch, and yaw, then evaluate the resulting model on the fruit-fly evasive maneuver. If error does not fall below the reported 6.7–15.4%—or if a robotic flapping wing in forward flight at advance ratio near 1.1 shows force RMSE no better than the conventional model—the claim that the three mechanisms capture the body-motion effects would be contradicted.","tokens_in":33649,"feed_emoji":"🐝","tokens_out":5208,"duration_ms":56548,"temperature":0.7,"pith_summary":"The paper aims to show that the quasi-steady model used to estimate aerodynamic forces on flapping wings can be made much more accurate outside hovering flight by adding three mechanisms that were qualitatively known but never formulated: the advance-ratio effect, the spanwise kinematic velocity effect, and the rotational Wagner effect. Using a sparse-regression search over roughly 5,000 candidate kinematic functions, with CFD force data for hawkmoth and fruit fly wings as ground truth, the authors derive explicit expressions for the three mechanisms and fold them into a single blade-element model. They report that the augmented model predicts instantaneous wing-normal forces with 11.0–16.0% error for hawkmoth forward flight and 6.7–15.4% error for fruit fly maneuvering flight, down from about 29% for the conventional model. The result matters because an accurate, fast quasi-steady model would let researchers scan insect flight strategies and design flapping robots without expensive CFD.","feed_headline":"Three added mechanisms cut flapping-wing force error from 29% to 7-16%","feed_subtitle":"A sparse-regression rewrite of the quasi-steady model adds advance-ratio, spanwise-flow, and rotational-Wagner terms.","key_machinery":"A blade-element quasi-steady model augmented by three formulated mechanisms. The load-bearing device is the decomposition of the wing surface velocity into body-velocity, flapping-velocity, and cross terms for both chordwise and spanwise blade elements, plus a Wagner-style delay term for rotational circulation. The sparse-regression selection from a library of 5,548 candidate functions is what identifies which of these terms the conventional model was missing.","core_discovery":"The central claim is that the failure of conventional quasi-steady models in forward and maneuvering flight is not an inherent limit of the quasi-steady assumption, but a set of three missing kinematic terms. The authors formulate (1) an advance-ratio effect in which translational lift and drag are decomposed into body-velocity, flapping-velocity, and coupling terms with separate coefficients; (2) a spanwise kinematic velocity effect in which blade elements are taken along the span, so the spanwise component of air velocity contributes translational and rotational forces; and (3) a rotational Wagner effect, in which the decay and growth of rotational circulation lag the wing's acceleration,","pith_inferences":["The same discovery pipeline could be reused to formulate additional missing terms—wing flexibility, wing-wing interaction, ground effect—by expanding the training data and candidate library; the paper notes the dataset would grow dramatically, so this is a testable extension rather than a demonstrated result.","Because the model is a linear combination of interpretable kinematic terms, it could be inverted to ask which wing kinematic adjustments produce a desired force change, giving a fast tool for flight-control and neurophysiology hypotheses.","A direct test of the weakest assumption would be to include body-rotation-induced velocity patterns in training data; if fruit-fly turn errors then fall further, the current model's maneuvering accuracy is partly lucky rather than fully mechanistic.","The vortex-breakdown residual suggests a hybrid design: quasi-steady model for attached-flow phases plus a stochastic or flow-state correction for breakdown episodes, rather than a purely kinematic quasi-steady model."],"forward_implications":["Force estimation errors drop from roughly 29% to 7–16% across hovering, forward, and maneuvering flight for hawkmoth and fruit fly, so the quasi-steady model becomes a fast substitute for CFD in parameter sweeps.","The advance-ratio decomposition means lift and drag coefficients are no longer constants: they vary with flight speed, changing how thrust and vertical force are predicted in forward flight.","Spanwise flow contributes over 10% of vertical force at several hawkmoth flight speeds, implying that spanwise blade elements should be included in future quasi-steady analyses.","The rotational Wagner effect adds a force term during stroke acceleration that is particularly visible in fruit-fly flight, extending the Wagner correction from translation to rotation.","Residual errors concentrate in upstroke episodes of leading-edge vortex breakdown; the model makes clear that flow-instability forces and wake capture remain outside quasi-steady reach."],"supporting_citations":[{"why":"Supplies the sparse-regression method used to rank and select candidate kinematic functions.","marker":"[51]"},{"why":"Provides the translational Wagner-effect model that the paper extends to rotational circulation.","marker":"[42]"},{"why":"Prior advance-ratio formulation that the algorithm rediscovers and reincorporates into the new model.","marker":"[75]"},{"why":"Gives qualitative evidence for spanwise lateral inflow effects that the paper formulates quantitatively.","marker":"[76]"},{"why":"Supplies the CFD-informed quasi-steady model and lift-coefficient form used as the conventional baseline.","marker":"[31]"},{"why":"Provides a prior CFD data-driven forward-flight quasi-steady model that the new model improves on.","marker":"[35]"},{"why":"Supplies fruit-fly evasive-maneuver kinematics used as validation data for the maneuvering case.","marker":"[3]"},{"why":"Supplies hawkmoth hovering and forward-flight kinematics used as validation data.","marker":"[50]"}],"fun_headline_variants":["Three new terms cut flapping-wing model error from 29% to 7-16%","Data-driven discovery pinpoints three missing terms to fix flight model","Quasi-steady flight model improved by adding three forgotten effects","Error drops to 7-16% after adding advance-ratio, spanwise, Wagner terms","Flapping-wing model gets data-driven fix: error drops from 29% to 7-16%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The training data set body angular velocity to zero, assuming wing-angle variations capture the influence of body rotation, while the fruit-fly test data include strong body turns; if that assumption fails, the reported accuracy on maneuvering flight is overstated.","fun_headline_variants_meta":{"raw":{"variants":["Three new terms cut flapping-wing model error from 29% to 7-16%","Data-driven discovery pinpoints three missing terms to fix flight model","Quasi-steady flight model improved by adding three forgotten effects","Error drops to 7-16% after adding advance-ratio, spanwise, Wagner terms","Flapping-wing model gets data-driven fix: error drops from 29% to 7-16%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":2996,"prompt_tokens":736,"completion_tokens":2260,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":2151}},"tokens_in":480,"tokens_out":2260,"duration_ms":20538,"temperature":1.0,"reasoning_tokens":2151,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:16:29.504246+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same sparse-regression pipeline on data that include nonzero body roll, pitch, and yaw, then evaluate the resulting model on the fruit-fly evasive maneuver. If error does not fall below the reported 6.7–15.4%—or if a robotic flapping wing in forward flight at advance ratio near 1.1 shows force RMSE no better than the conventional model—the claim that the three mechanisms capture the body-motion effects would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sparse-regression method used to rank and select candidate kinematic functions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior advance-ratio formulation that the algorithm rediscovers and reincorporates into the new model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives qualitative evidence for spanwise lateral inflow effects that the paper formulates quantitatively."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies hawkmoth hovering and forward-flight kinematics used as validation data."}],"review_version":1}