{"id":"f45d023b-3653-4bcf-bd9c-e45b42407459","arxiv_id":"2608.10309","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A machine-learned surrogate of urban wind fields can be used to plan drone trajectories that avoid turbulent wake regions, and was tested with a real quadrotor in a wind tunnel.","lead":"A team trained a U-Net model to predict wind flow and turbulence around city-like obstacles, then used those predictions to plan safer drone routes through a wind tunnel. The flow-informed routes reduced unwanted drift in one scenario, but needed manual tuning in a second scenario before beating a wind-ignoring baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scenario 2's safety advantage is established only after re-weighting the flight-challenge metric on the same outcome data; a prospective parameter-free test is needed.","rationale":"I agree with the reader that the flight-challenge metric and its weights are the weakest element, but the sharper failure is procedural: the metric's weights are selected after seeing the very outcome they are used to explain. This selection makes the Scenario 2 'improvement' non-identifiable as evidence. The paper is otherwise careful: the surrogate test error is reported, the solver and data pipeline are described, and Scenario 1 gives independent support with a large effect. I would not reject the work; the framework is promising and the authors disclose the limitation. However, the central abstract claim should remain conditional on a prospective validation that fixes F before comparing trajectories. The three-trial design compounds the issue: even if the weights had been fixed, a 0.6 cm mean difference could be noise. The proposed re-test is feasible in the same facility and would settle whether the headline claim survives. I therefore keep the reader's conditional verdict and recommend no change.","tokens_in":9309,"tokens_out":6803,"duration_ms":75139,"concrete_test":"Prospective re-test of Scenario 2: freeze the flight-challenge weights to the pre-tuned CFD-derived values (or, preferably, to a protocol specified before any flight data are seen), then run at least 10 trials each for surrogate-informed, CFD-informed, and wind-naive trajectories, and report per-trial mean displacement with a paired permutation test for surrogate versus naive. If the surrogate-informed mean is not significantly lower (e.g., paired p < 0.05) or the unadjusted-weights version again loses to naive, the headline claim of reduced displacement is not established. Re-running with pre-registered weights is the single check that separates genuine improvement from parameter selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the flow-informed pipeline reduces vehicle displacement relative to a wind-naive baseline. In Scenario 1 that claim is credible, but Scenario 2 is the case that distinguishes the surrogate from a trivial baseline, and there the default metric fails: with the weights set manually from CFD fields (Eq. 1, Section 3.1.2), the surrogate-informed path crosses a near-wake shear layer and has mean displacement 6.7 cm versus 5.5 cm for the naive path. The reported 4.9 cm advantage is obtained only after increasing w2 (the velocity-gradient weight) in response to this failure. Because the same experiment that motivated the re-weighting is then cited as evidence for the improvement, the comparison is not a prospective test; it is a post hoc fit. The Discussion acknowledges this, conceding that tuning 'does not fix the underlying cause' and that unforeseen issues may remain. With three trials per condition and no error bars or significance tests, the claimed reduction in Scenario 2 could easily reflect weight selection plus noise. The load-bearing condition is therefore not only that F is physically reasonable, but also that its parameters are fixed before the comparative claim is evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end pipeline for wind-aware flight planning: a U-Net surrogate trained on 3392 time-averaged Lattice-Boltzmann flow fields predicts three-dimensional velocity and TKE from building geometry and incident wind; a hand-specified 'flight challenge' scalar field (Eq. 1) combines normalized velocity-gradient magnitude, normalized TKE, and obstacle proximity with weights w0-w4; Dijkstra's algorithm then finds a minimum-cost path. The system is demonstrated in a fan-array wind tunnel on two model urban geometries, comparing surrogate-informed, CFD-informed, and flow-naive trajectories, with three trials per condition. In Scenario 1 the CFD- and surrogate-informed paths have mean displacements of 7.7 cm and 7.4 cm versus 18.8 cm for the naive path. In Scenario 2 the default surrogate-informed path has mean displacement 6.7 cm, worse than the naive path's 5.5 cm; after increasing the velocity-gradient weight w2 in Eq. (1), the surrogate-informed path achieves 4.9 cm. The authors conclude that the flow-informed approach reduces undesired vehicle displacement and improves flight stability.","tokens_in":9519,"tokens_out":5311,"duration_ms":54797,"significance":"If the central claim were established by a prospective, statistically adequate experiment, this would be a meaningful practical demonstration of machine-learning-based wind-aware flight planning in a physical urban-like environment. The paper's strengths include the full-stack validation, the use of a 3D U-Net with a quantified holdout error of 8.71%, and explicit comparison against both CFD-informed and flow-naive baselines. However, the Scenario 2 default result directly contradicts the abstract's claim, the reported improvement is obtained only after post hoc adjustment of the flight-challenge weights, and no uncertainty quantification is provided for any of the displacement means. As presented, the load-bearing evidence is therefore not yet convincing; a prospective evaluation with fixed or cross-validated weights and quantitative error bars is needed. The work is honest about its limitations, but those limitations currently affect the main empirical claim.","major_comments":[{"comment":"With the weights set as described in Section 3.1.2, the surrogate-informed trajectory in Scenario 2 has a mean displacement of 6.7 cm, which is worse than the flow-naive path's 5.5 cm; this contradicts the abstract's claim that the flow-informed approach reduces undesired vehicle displacement. The reported 4.9 cm value is obtained only after increasing w2 following the observed failure, so the same experimental outcome is used both to motivate the re-weighting and to measure the improvement. This is a post hoc fit rather than a prospective comparison. The authors should provide a prospective test with weights fixed before the flight experiments, or a cross-validation in which weights are selected on a separate calibration scenario and then evaluated on held-out flight data.","section":"Section 3.1.2, Eq. (1)"},{"comment":"Only three trials per condition are reported, with no standard deviations, confidence intervals, or significance tests. In Scenario 2 the relevant differences are small (5.4 vs 5.5 cm and 4.9 vs 5.5 cm) relative to the spread visible in Figure 8a, so the claimed advantage could easily be sampling noise. The authors should report per-trial displacement values, error bars on all mean displacements, and an appropriate statistical comparison such as a bootstrap or permutation test for the pairwise differences.","section":"Section 2.2, Section 3.1"},{"comment":"The paper attributes the surrogate-informed path's wake crossing to 'overly smooth predictions' and 'artificially inflated flight challenge values' downstream, but no quantitative link is established between the surrogate's prediction error and the resulting path. The difference map in Figure 8c is qualitative, and the text later concedes that tuning 'does not fix the underlying cause.' Without an analysis of which flow-field error component (velocity magnitude, velocity gradient, TKE, or their spatial location) caused the path to cross the shear layer, the conclusion that the surrogate model is the main limiting factor is not supported; the sensitivity of the planned path to plausible surrogate errors should be checked directly.","section":"Section 3.1.2, Section 4"},{"comment":"The flight challenge metric depends on five manually set weights w0-w4, but the paper does not report the actual weight values, the range of values explored, or a sensitivity analysis. Since the weights are adjusted after observing experimental outcomes, it is impossible to determine how much of the reported improvement reflects the metric formulation rather than the flight planner or the surrogate. The authors should state the exact weights used for each reported trajectory and provide a sensitivity analysis over a plausible weight range showing that the Scenario 1 and Scenario 2 conclusions are robust to reasonable variations in the weights.","section":"Section 2.1, Eq. (1)"}],"minor_comments":[{"comment":"The abstract and Figure 1 caption refer to 'time-resolved flow fields,' but the surrogate is trained and evaluated on time-averaged velocity and TKE fields; please use consistent terminology throughout.","section":"Abstract, Figure 1"},{"comment":"There is a typo 'UA V' in the Introduction; also, the hyphenation of 'Lattice-Boltzmann' is inconsistent in a few places.","section":"Introduction"},{"comment":"The accumulated flight challenge metric in Figure 6 is computed from the CFD-generated field even for the surrogate-informed trajectory; please define how the accumulation is performed and clarify why the surrogate-informed trajectory is evaluated on the CFD field rather than the surrogate field.","section":"Figure 6"},{"comment":"The normalization of the obstacle proximity D is not specified. Please state how the Gaussian-convolved distance field is normalized so that the weight w3 and w4 terms are dimensionally and numerically comparable to the normalized flow terms.","section":"Section 2.1, Eq. (1)"},{"comment":"The claim of being the 'first real-world demonstration' of wind-aware flight planning should be tempered: Habib et al. [2026] report micro-UAV experiments with wind-adaptive planning around box fans, and Achermann et al. [2024] field-tested a volumetric wind predictor aboard a UAV. The novelty statement should distinguish the present large-fan-array urban-obstacle setting from these prior experimental demonstrations.","section":"Section 1, Related Work"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty framing is somewhat vulnerable given the existing experimental works by Habib et al. and Achermann et al. The 'first practical demonstration' phrasing in the abstract and Introduction should be narrowed or supported with a clearer comparison to those prior experiments. Also, the data availability statement ('available upon request') is weak for a methods-oriented paper; if the journal has a reproducibility policy, the authors should be encouraged to archive code, trained model weights, and flight logs. The core issue for the decision is the Scenario 2 post hoc weight tuning; if the authors can supply a prospective or cross-validated evaluation with error bars, the paper would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Paper is worth a serious look, but the headline claim overreaches. The actual new thing is an integrated pipeline: a 3D U-Net surrogate for urban flow, a hand-built flight challenge field, Dijkstra planning, and real quadrotor tests in a fan-array wind tunnel. Scenario 1 is clean: both surrogate- and CFD-informed paths avoid wakes and cut mean displacement roughly in half (7.4 cm vs 18.8 cm). The authors also deserve credit for reporting Scenario 2's failure openly and discussing the smoothing problem in learned surrogates.\n\nThe soft spot is Scenario 2. With the hand-tuned weights set from CFD fields, the surrogate-informed path is worse than the naive baseline (6.7 cm vs 5.5 cm). The reported 4.9 cm comes only after increasing w2 in response to that failure. That is post hoc tuning, and the paper is candid about it. It means the safety improvement claim is not yet supported as a general statement. Three trials per condition and no variance or significance reporting makes the 4.9 vs 5.5 difference impossible to evaluate. Also, CFD-informed performance in Scenario 2 is essentially tied with the naive baseline (5.4 vs 5.5), which suggests the challenge metric itself may not be capturing the hazard in that configuration.\n\nThe literature context is handled honestly. The prior works are cited, including the 2D methods, WindSeer field tests, and the WESPR box-fan hardware test, so the 'first real-world demonstration' framing is a bit strong. Data and code are 'available upon request,' which is not reproducible.\n\nVerdict: this deserves peer review, not a desk reject. The pipeline is sensible, Scenario 1 is a real result, and the discussion of limitations is honest. A referee should require the following: a prospective test with weights fixed before the experiment, more trials or confidence intervals, a comparison of the tuned weights to the original ones, and public release of the surrogate model and flight logs. If those are addressed, the paper could be a useful reference for urban AAM planning.","headline":"Integrated surrogate-plus-planner pipeline with a clean Scenario 1 hardware demonstration, but the Scenario 2 safety gain rests on post hoc weight tuning.","tokens_in":10124,"tokens_out":2909,"would_cite":true,"duration_ms":27285,"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 machine-learning surrogate trained on CFD can predict urban wind fields fast enough that a cost-minimizing pathfinder can route drones around hazardous flow, and wind-tunnel flights show these flow-informed paths displace less and fly…","keywords":["advanced air mobility","urban wind","flight planning","surrogate model","U-Net","turbulent kinetic energy","flight challenge metric","wind tunnel validation"],"falsifier":"A concrete test is to fly several different planned paths through a region whose surrogate-predicted flight challenge is low but whose CFD field shows a sharp shear layer, then measure displacement and angular velocity; if a low-challenge path consistently produces large deviations, the metric's ranking is wrong. The paper's Scenario 2 is already an instance: before weight tuning, the surrogate-informed path crossed the near wake and displaced 6.7 cm, worse than the 5.5 cm naive path. A systematic version would scan a grid of metric weights and check whether the lowest-displacement path aligns with the lowest predicted challenge across multiple geometries.","tokens_in":9125,"feed_emoji":"🌬️","tokens_out":6681,"duration_ms":58916,"temperature":0.7,"pith_summary":"This paper argues that a flight planner for small aerial vehicles can be made safer by first predicting the wind field around buildings with a machine-learning surrogate, converting that prediction into a scalar 'flight challenge' map, and then routing through the map with a cost-minimizing pathfinder. The claim is supported by wind-tunnel flights: paths planned from wind knowledge kept the test quadrotor closer to its target and steadier than paths planned without wind knowledge. If correct, the result suggests that real-time wind-aware routing in complex urban terrain is within reach using only building geometry and an incident wind reading.","feed_headline":"Wind-aware flight paths beat wind-naive routes in tunnel tests","feed_subtitle":"A surrogate-planned quadrotor route held 7.4 cm mean displacement, versus 18.8 cm for a wind-ignoring path.","key_machinery":"The central object is the flight challenge metric $F(x,y,z) = w_1(w_2\\tilde{G} + \\tilde{k})^2 + w_3 D + w_4 D(w_2\\tilde{G} + \\tilde{k}) + w_0$, where $\\tilde{G}$ and $\\tilde{k}$ are the min-max normalized velocity-gradient magnitude and turbulent kinetic energy, $D$ is the Gaussian obstacle-proximity penalty, and the $w$ terms are hand-set weights. This scalar field converts a flow prediction into a cost map that encodes shear layers, turbulence, and structure proximity, and it is what makes the pathfinder produce wind-aware routes. The second load-bearing component is a U-Net surrogate that maps a signed-distance encoding of building geometry plus incident wind speed to volumetric three-component velocity and TKE predictions, with a reported test error of 8.71 percent normalized $\\ell^2$. A Dijkstra pathfinder then minimizes accumulated $F$ along the route.","core_discovery":"The paper's central claim is that a flow-informed trajectory generated from a surrogate-predicted wind field is safer and more stable than a flow-naive shortest path. In the first obstacle geometry, the surrogate-informed and CFD-informed paths produced mean displacements of 7.4 cm and 7.7 cm, respectively, while the naive path averaged 18.8 cm and showed trial-to-trial scatter and collisions. In the second geometry, the unadjusted surrogate-informed path passed through a near-wake shear layer and underperformed the naive baseline (6.7 cm versus 5.5 cm), but after increasing the velocity-gradient weight in the flight challenge metric, the surrogate-informed path achieved 4.9 cm mean displacement, beating the 5.5 cm naive path. The authors attribute the need for tuning to oversmoothing in the surrogate's predicted fields.","pith_inferences":["The oversmoothing failure seen in Scenario 2 suggests a testable comparison: adversarially trained or generative surrogates that sharpen wakes should make the fixed-weight metric rank shear layers correctly, removing the manual tuning step.","Because the challenge metric is vehicle-agnostic, a natural extension is to replace the fixed weights with vehicle-specific disturbance-response models, so a heavy, slow aircraft and a light, fast one get different routes through the same wind field.","The same pipeline could be inverted to define conservative 'no-fly' zones for a fleet with minimum performance guarantees rather than a single trajectory, which would be a direct path to certification-style safety envelopes.","A stronger experimental test would fly the same geometries at several incident wind speeds and with the surrogate retrained on higher-resolution data, to see whether the Scenario 2 shear-layer mis-ranking disappears as prediction sharpness improves."],"forward_implications":["Wind-aware planning reduces mean trajectory displacement by more than a factor of two relative to a wind-naive baseline in the tested urban-like geometry.","The flight challenge metric's correlation with accumulated displacement and angular velocity supports its use as a planning cost rather than only a post-flight diagnostic.","A surrogate trained once on CFD can be reused for new geometries and wind speeds without new full-scale simulations, keeping planning fast enough for near-real-time use.","The metric's weights need case-by-case tuning when the surrogate output is smoother than CFD; the paper shows one such adjustment recovering performance in the second scenario."],"supporting_citations":[{"why":"Establishes the prior approach of learning local urban wind fields from sensing and planning paths from them, which this work extends to 3D and experimental validation.","marker":"[Folk et al., 2024]"},{"why":"Provides the wind-aware motion-planning baseline this system builds on, though that planner accessed simulated data directly.","marker":"[Folk et al., 2025]"},{"why":"Motivates the U-Net surrogate architecture with skip connections for reconstructing fine-scale flow features.","marker":"[Renn et al., 2026]"},{"why":"Supplies the GPU-accelerated Lattice-Boltzmann solver used to generate the 3,392 training flow fields.","marker":"[Ataei and Salehipour, 2024]"},{"why":"Defines the atmospheric boundary layer inlet profile used in the CFD training data.","marker":"[Richards and Hoxey, 1993]"},{"why":"Provides the multi-quadrotor control and localization stack used to fly the test vehicle along planned paths.","marker":"[Preiss* et al., 2017]"},{"why":"Documents the oversmoothing shortfall of deep surrogate models and proposes adversarial training as a fix, which the paper invokes to explain Scenario 2.","marker":"[Oommen et al., 2026]"},{"why":"Supplies the Dijkstra pathfinding implementation used to minimize flight challenge along the route.","marker":"[Veiga-Piñeiro et al., 2025]"}],"fun_headline_variants":["Surrogate wind model steers drone path, reducing drift in tunnel tests","ML-predicted wind fields pilot safer urban drone routes in wind tunnel","Flow-informed drone path beats wind-naive route in urban canyon trial","Wind-aware flight planner outperforms wind-naive in wind tunnel tests","Surrogate-based wind planning reduces drone displacement in urban wind tunnel"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-tuned flight challenge metric, evaluated on surrogate-predicted time-averaged wind velocity and turbulence fields, correctly ranks aerodynamic hazard for the test vehicle; if the weights or the surrogate's smoothed fields misrank a dangerous shear zone as benign, the planned 'safe' path is not actually safe.","fun_headline_variants_meta":{"raw":{"variants":["Surrogate wind model steers drone path, reducing drift in tunnel tests","ML-predicted wind fields pilot safer urban drone routes in wind tunnel","Flow-informed drone path beats wind-naive route in urban canyon trial","Wind-aware flight planner outperforms wind-naive in wind tunnel tests","Surrogate-based wind planning reduces drone displacement in urban wind tunnel"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000791,"raw_usage":{"total_tokens":3468,"prompt_tokens":912,"completion_tokens":2556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":528,"tokens_out":2556,"duration_ms":18189,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:10:13.017362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is to fly several different planned paths through a region whose surrogate-predicted flight challenge is low but whose CFD field shows a sharp shear layer, then measure displacement and angular velocity; if a low-challenge path consistently produces large deviations, the metric's ranking is wrong. The paper's Scenario 2 is already an instance: before weight tuning, the surrogate-informed path crossed the near wake and displaced 6.7 cm, worse than the 5.5 cm naive path. A systematic version would scan a grid of metric weights and check whether the lowest-displacement path aligns with the lowest predicted challenge across multiple geometries.","supporting_citations":[{"cited_title":"Learning local urban wind flow fields from range sensing","cited_arxiv_id":null,"evidence_quote":"Establishes the prior approach of learning local urban wind fields from sensing and planning paths from them, which this work extends to 3D and experimental validation."},{"cited_title":"Xlb: A differentiable massively parallel lattice boltzmann library in python","cited_arxiv_id":null,"evidence_quote":"Supplies the GPU-accelerated Lattice-Boltzmann solver used to generate the 3,392 training flow fields."},{"cited_title":"Appropriate boundary conditions for computational wind engineering models using the k- turbulence model","cited_arxiv_id":null,"evidence_quote":"Defines the atmospheric boundary layer inlet profile used in the CFD training data."},{"cited_title":"Learning turbulent flows with generative models for super resolution and sparse flow reconstruction","cited_arxiv_id":null,"evidence_quote":"Documents the oversmoothing shortfall of deep surrogate models and proposes adversarial training as a fix, which the paper invokes to explain Scenario 2."},{"cited_title":"Hybrid cfd-deep learning approach for urban wind flow predictions and risk-aware uav path planning","cited_arxiv_id":null,"evidence_quote":"Supplies the Dijkstra pathfinding implementation used to minimize flight challenge along the route."}],"review_version":1}