{"id":"e7970447-1012-445f-973b-3edcd555d236","arxiv_id":"2607.26683","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Convolutional information-theoretic learning decomposes extreme vortex-gust airfoil flows at Re=5000 into structures causally informative for future lift and scale-dependent energy transfer.","lead":"A machine-learning method pulls out only the swirl patterns in a gust hitting a wing that actually matter for future lift or energy transfer. It gives aerodynamicists a causal, snapshot-based way to study extreme, multiscale gust encounters without needing the full governing equations.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Causal extraction claim hinges on untested equivalence between MI-sufficient subsets (network given future λ) and dynamical influence","rationale":"The reader already isolates the precise soft spot: MI maximization under the non-negative 3D-CNN + bijective architecture with hand-chosen β/Δt yields association, not demonstrated physical causality. The manuscript’s own §4 notes that interventional tests would be needed. Force-element agreement is only qualitative and same-time, so it does not close the gap. No stronger internal inconsistency appears; the method is a coherent extension of the authors’ prior framework and the pre-/post-impingement structural observations are clearly reported. Verdict therefore stays CONDITIONAL pending quantitative predictive-skill or intervention checks and public artifacts; no upgrade or downgrade is warranted.","tokens_in":14950,"tokens_out":524,"duration_ms":40176,"concrete_test":"On held-out gust cases, train simple regressors (e.g., a small CNN or kernel ridge) to predict CL(t+Δt) from q_I alone, from q_R alone, and from full q; report R² or mutual information. If R²(q_I) is not near R²(q) while R²(q_R) remains appreciable (or if random large-scale masks of comparable ||q_I|| match the skill), the informativeness/causality claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim treats q_I from Eqs. 2–3 as the structures “causally responsible” for future λ (CL or T). The extractor F is conditioned on the realized future scalar λ(t+Δt) itself, and the loss only minimizes ||q−q_I||² plus residual independence; H(λ|q_I)=0 is imposed by non-negative weights and bijective activations rather than by an explicit predictive objective. Among all subsets informationally sufficient for the observed λ, the L2 term therefore selects the largest (in field norm) such subset. Force-element comparison (Eq. 7) is instantaneous and only qualitative (isosurface overlap in Figs. 4–5), so it does not validate the future-oriented or causal reading. Without a check that q_I alone retains predictive skill for λ while q_R does not, or that intervening on q_I changes λ, the structures may be statistically associated large-scale features preferred by the inductive bias rather than the dynamical drivers of the target.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript applies convolutional information-theoretic mode decomposition to LES of extreme vortex-gust–airfoil interactions at Re=5000 (G=±{2,5}). A 3D CNN (deep sigmoidal flow with non-negative weights) decomposes each Q-criterion snapshot into informative and residual fields by maximizing mutual information with a future scalar λ while enforcing residual independence (Eqs. 1–3). Two choices of λ are examined: lift coefficient CL and domain-integrated scale-dependent energy transfer T. Lift-targeted modes are compared visually to instantaneous volume force elements; energy-targeted modes are compared to lift-targeted modes via a spatial cosine-similarity time series that rises from ~0.4 pre-impingement to ~0.8 after. The authors conclude that the method selectively extracts structures responsible for the chosen future physics and can support causal, data-driven study of transient aerodynamics.","tokens_in":15223,"tokens_out":1477,"duration_ms":41078,"significance":"If the extracted fields are accepted as physically meaningful, the work usefully extends information-theoretic decomposition from prior 2D/separated-wake settings to three-dimensional turbulent extreme-gust encounters and, importantly, demonstrates target-dependence across two distinct aerodynamic mechanisms (force vs inter-scale transfer). Strengths include a well-documented LES setup, multi-gust-ratio coverage, an external physical check against force elements (independent of the MI objective), Q–R topology diagnostics, and an explicit cross-target comparison. The framework is in principle applicable to experimental data where auxiliary potentials are unavailable. These are genuine contributions to data-driven analysis of extreme aerodynamics, provided the causal reading is either validated or carefully scoped.","major_comments":[{"comment":"§2.1, Eqs. (2)–(3) and the abstract/title framing: the extractor F is conditioned on the realized future scalar λ(t+Δt), and the loss only balances reconstruction against residual independence; H(λ|q_I)=0 is imposed by non-negative weights and bijective activations rather than by an explicit out-of-sample predictive objective. Among MI-sufficient subsets, the L2 term therefore favors large-norm features. The manuscript equates these fields with structures “causally responsible” for future λ, yet provides no check that q_I alone retains predictive skill for λ while q_R does not, nor any interventional test. The force-element comparison (Eq. 7) is instantaneous and only qualitative (isosurface overlap in Figs. 4–5), so it does not validate the future-oriented causal claim. Either (i) add a simple hold-out prediction or ablation (forecast λ from q_I vs q_R) or (ii) systematically replace “c","section":"§2.1, Eqs. (2)–(3); Abstract; §3.1"},{"comment":"§3.1, Figs. 4–5 and Eq. (7): agreement with lift elements is asserted solely by visual isosurface correspondence. Because force elements diagnose instantaneous force generation while q_I is defined with respect to CL(t+Δt), a quantitative spatial metric (e.g., cosine similarity or overlap of thresholded supports between |Le| and q_I as functions of Δt and t) is needed to make the comparison falsifiable and to show that the time-delay effect in Fig. 6 is not merely a change in visualization. Without it, the external physical check remains impressionistic.","section":"§3.1, Figs. 4–5, Eq. (7)"},{"comment":"§3.2, Eq. (13) and Fig. 9: the claimed analogy between lift- and energy-targeted structures rests on a single cosine-similarity time series for G=2 (rising to ~0.8). The manuscript does not report the same metric for G=5 or the negative-gust cases, nor sensitivity to the band-pass scales defining T (σ_max, σ1) or to Δt. Given that free parameters (β via L-curve, Δt, filter scales, Q_th) are numerous, at least one additional gust ratio and a brief sensitivity check on the similarity curve are required before the cross-mechanism “analogy after impingement” can be treated as a robust result rather than a single-case observation.","section":"§3.2, Eq. (13), Fig. 9"}],"minor_comments":[{"comment":"The L-curve used to fix β is cited (§2.1) but never shown; a brief inset or appendix panel would make the regularization choice reproducible.","section":"§2.1"},{"comment":"POD citation is broken in the text (“proper orthogonal decomposition [?, POD;]]lumley1967structure”). Restore a proper Lumley/Sirovich reference.","section":"§3.1"},{"comment":"Fig. 2 architecture: filter sizes and channel counts are given, but training set size (number of snapshots × cases), batch size, learning rate, and early-stopping patience are not stated. Add a short paragraph or table for reproducibility.","section":"§2.1, Fig. 2"},{"comment":"Q_th = 0.1 is used uniformly for visualization; state whether informative/residual fields are thresholded identically and whether conclusions change under modest threshold variation.","section":"Figs. 4–6, 8–9"},{"comment":"Acknowledgements list R.A. and Q.L. who are not on the author line; clarify contributions or authorship.","section":"Acknowledgements"},{"comment":"Minor prose: “asextreme” (p.1), “that decomposes” fragment in §2.1 opening, and “R.A.” vs author list consistency. A careful copy-edit pass is needed.","section":"Introduction; §2.1"}],"recommendation":"major_revision","confidential_remarks":"The methodological core is closely continuous with the authors’ own recent JFM/AIAA papers on convolutional causal learning; the main novelty is the Re=5000 extreme-gust application and the dual-target (CL vs T) comparison. That is appropriate for a fluids journal if the causal language is disciplined and the quantitative gaps above are closed. I do not see a soundness error that would warrant rejection, but accepting the manuscript with the present causal framing and purely qualitative external checks would set a low bar for “causal extraction” claims in the area."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new material here is the application, not the method. They take the convolutional informative/non-informative decomposition (Arranz–Lozano-Durán plus their own 2D/lower-Re papers) and run it on 3D LES of extreme vortex-gust–airfoil interactions at Re=5000 across four gust ratios, with two future targets: lift and scale-dependent energy transfer. That dual-target comparison is the real addition.\n\nWhat they do well is the physical side of the application. Pre-impingement the lift-targeted modes pick cores; after massive separation they also keep the emerging shear layers. That lines up qualitatively with instantaneous force-element maps (Figs. 4–5), and the Q–R topology (Fig. 7) is a clean extra check. The cosine-similarity time series between lift- and energy-targeted fields (rising from ~0.4 to ~0.8 after impingement) is a concrete, new observation on this configuration. LES setup, scale-decomposition formulas, and network architecture are stated clearly enough to follow.\n\nThe soft spot is exactly the one the stress-test flags, and it is real but not fatal. The extractor is conditioned on the realized future scalar λ(t+Δt); the loss is reconstruction plus residual independence, with H(λ|q_I)=0 enforced by non-negative weights and bijective activations rather than by an explicit predictive or interventional test. So “causally responsible” is stronger language than the evidence supports. Force-element agreement is instantaneous and visual only. They themselves note in the outlook that intervention would be the proper check. Free parameters (β, Δt, architecture, scales) are acknowledged via L-curve and sweeps, but code/data are not released. Circularity is modest because the force-element comparison is external to the MI objective.\n\nThis is for people already working extreme aerodynamics or data-driven structure extraction who want a practical way to isolate target-specific features from 3D gust snapshots. It is not a foundational theory paper. I would send it to referees; the application and the dual-target result are worth the community’s time, provided the causal wording is tightened and some quantitative predictive check is added. I would cite the cosine-similarity finding and the shear-layer observation if I were writing on gust–airfoil interaction.","headline":"Solid extension of the authors’ causal-decomposition tool to 3D extreme gusts at Re=5000, with useful dual-target physics; the “causal” label still outruns the evidence.","tokens_in":15864,"tokens_out":588,"would_cite":true,"duration_ms":12792,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A convolutional information-theoretic network splits extreme gust–airfoil flow snapshots into the vortical pieces that carry future lift or energy transfer and the residual that does not.","keywords":["extreme aerodynamics","vortex gust","informative mode decomposition","information-theoretic learning","lift coefficient","scale-dependent energy transfer","convolutional neural networks","turbulent wake"],"falsifier":"In a new simulation, locally suppress or perturb only the extracted informative structures and measure whether the subsequent lift (or integrated energy transfer) changes far more than when the residual structures are perturbed by the same amplitude; comparable effects would falsify the causal claim.","tokens_in":15780,"feed_emoji":"🌀","tokens_out":902,"duration_ms":42533,"temperature":0.7,"pith_summary":"Extreme vortex gusts hitting an airfoil at Reynolds number 5000 produce fast, multiscale wakes in which it is hard to tell which structures actually set later aerodynamic forces or energy exchange. This paper trains a three-dimensional convolutional network with an information-theoretic objective so that any single Q-criterion snapshot is decomposed into an informative part that maximises mutual information with a chosen future scalar and a residual that is independent of that part. When the scalar is future lift, the network first highlights large vortex cores and, after massive separation, also the emerging shear layers, in qualitative agreement with instantaneous force-element maps. When the scalar is scale-dependent energy transfer, the extracted structures differ from the lift-based ones before impingement yet become similar afterward. The method therefore lets a researcher ask, from data alone, which pieces of a transient turbulent field are tied to the physics of interest.","feed_headline":"ML isolates vortices that drive future lift in gust hits","feed_subtitle":"An information-theoretic net splits each snapshot into the pieces that set lift or energy transfer and the rest.","key_machinery":"Informative mode decomposition: a deep-sigmoidal-flow 3-D CNN that realises q = q_I + q_R by maximising mutual information I(λ; q_I) while enforcing I(q_R; q_I) = 0 through non-negative weights and bijective activations, with the future target λ and time lag Δt supplied as inputs.","core_discovery":"Convolutional information-theoretic learning decomposes a Q-criterion snapshot of an extreme vortex-gust–airfoil interaction into informative and residual components with respect to an arbitrary future target. For future lift the extracted modes are mainly vortex cores before impingement and additionally shear layers after separation, matching force-element locations; for scale-dependent energy transfer the modes are distinct early yet converge (spatial cosine similarity rising toward 0.8) after impingement, showing a shared structural basis across the two mechanisms.","pith_inferences":["Because only observables are required, the same extractor could be trained on multi-camera PIV of laboratory gust encounters and still flag the post-separation shear layers.","The post-impingement rise in similarity between lift and energy modes suggests a single set of structures whose suppression might reduce both force spikes and cascade activity.","Replacing the mutual-information loss with an interventional (do-operator) loss could further shrink the extracted set to the true causal skeleton."],"forward_implications":["Lift-driving structures can be read from experimental snapshots without constructing an auxiliary potential for force-element analysis.","The same snapshot can be re-decomposed for different targets (lift versus energy transfer) to compare mechanisms side by side.","Varying the time lag Δt reveals which structures matter on short versus longer horizons.","Spatial maps of informative modes can indicate where and when to actuate for gust-load control.","The framework applies in principle to any transient aerodynamic flow that supplies snapshots and a scalar target."],"fun_headline_variants":["ML isolates lift-driving vortices in extreme gust hits","Info net extracts vortices that set future lift","Causal ML splits gust flow into lift-relevant structures","Vortices tied to lift and energy transfer converge post-hit","Net finds shared structures across lift and energy mechanisms"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That maximising mutual information between a network-extracted field piece and a future scalar, under the stated architectural constraints, isolates structures that dynamically cause that scalar rather than merely correlating with it.","fun_headline_variants_meta":{"raw":{"variants":["ML isolates lift-driving vortices in extreme gust hits","Info net extracts vortices that set future lift","Causal ML splits gust flow into lift-relevant structures","Vortices tied to lift and energy transfer converge post-hit","Net finds shared structures across lift and energy mechanisms"]},"model":"grok-4.5","effort":"low","cost_usd":0.003597,"raw_usage":{"total_tokens":1202,"prompt_tokens":795,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":35968000,"prompt_tokens_details":{"text_tokens":795,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":348,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":795,"tokens_out":59,"duration_ms":6012,"temperature":1.0,"reasoning_tokens":348,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T00:05:53.915389+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"In a new simulation, locally suppress or perturb only the extracted informative structures and measure whether the subsequent lift (or integrated energy transfer) changes far more than when the residual structures are perturbed by the same amplitude; comparable effects would falsify the causal claim.","supporting_citations":[],"review_version":1}