{"id":"7d95be7e-17a6-4fee-9ad0-cec30902116a","arxiv_id":"2510.11600","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An offline-trained LSTM latent-dynamics model, SHAP-selected surface pressure sensors, and gradient-based MPC cut drag by 12.8% in a 2D truck-wake DNS at Re=500.","lead":"The authors built a closed-loop drag-reduction system for a simplified 2D truck wake: an LSTM-based latent-dynamics model trained offline on open-loop pressure data, with SHAP-selected surface sensors and a gradient-based model predictive controller. In a DNS at Reynolds number 500 using only four pressure sensors, the controller reduced drag by 12.8%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 12.8% drag reduction rests on one closed-loop DNS run in a regime the model is acknowledged to extrapolate; repeated trials and a periodic-forcing baseline are needed before this is a robust property of the framework.","rationale":"The reader's weakest assumption—that the closed-loop result depends on latent-model accuracy in an extrapolated regime, with a single simulation run bounding no risk—is exactly the load-bearing concern. The paper is transparent about this limitation, and the architecture (offline-trained LSTM latent dynamics, VICReg regularization, SHAP-based sensor pruning, differentiable MPC) is well described and technically coherent. The open-loop force-estimation accuracy (R² = 0.94 for Cd on the chirp test) and the qualitative physical consistency of the controlled wake give real support to the framework. However, the central quantitative claim is a single point estimate from one initial condition, and the section most relevant to that claim explicitly acknowledges model extrapolation. A single run cannot bound variability due to initial phase, model training seed, or the absence of a trivial periodic-forcing baseline. These omissions do not make the result false; they make it conditional. The reader's CONDITIONAL verdict is therefore appropriate, and no adjustment is needed.","tokens_in":15917,"tokens_out":9587,"duration_ms":94910,"concrete_test":"Repeat the closed-loop DNS with the same trained controller from at least 5–10 distinct initial conditions/phases (or with independently retrained LSTM models, if stochasticity is to be covered), computing mean Cd over a matched stationary interval such as t ∈ [20, 100] tc. In the same setup, also run an open-loop periodic-forcing baseline at the frequency and amplitude to which the MPC action converges. If the mean reductions are not consistently above baseline uncertainty and clearly larger than the open-loop periodic forcing, the 12.8% headline should be treated as a single-trajectory demonstration rather than a robust validation of the framework.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result is the single closed-loop DNS trajectory in Fig. 11: mean Cd drops from a baseline of 1.051 to 0.916 under MPC. The reliability of that number depends on the learned latent model being trustworthy exactly where the paper says it is not. In §3.3, the authors state that as the wake stabilizes, the model 'departs further from its interpolation region.' Since the MPC cost (Eqs. 13–15) is minimized over model-predicted Cd, a biased prediction in that extrapolated regime can select actions that are not optimal for the true plant. The recorded trajectory cannot by itself separate a genuinely robust controller from a favorable initial phase/transient. No repeated runs, no uncertainty quantification, and no open-loop periodic-forcing baseline are provided. The point estimate 1.051→0.916 is plausible and internally consistent, but it is not yet established as a stable property of the framework: the same controller could give a substantially different reduction from another initial condition, and the absence of a simple periodic-forcing comparison leaves open the possibility that the observed reduction is mostly due to actuation at the shedding frequency rather than to the MPC/latent-model machinery.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end data-driven active flow control framework that combines an offline-trained LSTM-based latent dynamics model with gradient-based model predictive control (MPC). The latent state is inferred from surface-mounted pressure sensors, and SHAP-based attribution is used to reduce the sensor set from 90 to four. The method is demonstrated in a 2D DNS of a simplified truck wake at Re=500, where the authors report a 12.8% drag reduction (Cd from 1.051 to 0.916) under closed-loop MPC. The drag reduction is directly measured in the DNS and the methodology is described in detail. However, the central demonstration rests on a single closed-loop trajectory, with no repeated runs, no uncertainty quantification, and no comparison against simpler open-loop forcing. The paper also acknowledges that the model is extrapolating in the closed-loop regime (§3.3), which raises a specific concern about the reliability of the MPC-optimized actions.","tokens_in":16226,"tokens_out":4553,"duration_ms":41987,"significance":"If the reported drag reduction is robust, this is a valuable contribution to data-driven flow control. The framework addresses practical deployment constraints—offline training, few non-intrusive surface sensors, and a differentiable latent model that makes MPC computationally tractable. The SHAP-based sensor selection is interpretable and the four-sensor ``slim'' encoder achieves force-prediction accuracy comparable to the full 90-sensor model on the test set (§3.2, Fig. 10). The 12.8% drag reduction is a direct DNS measurement, not an artifact of the model. The main limitation is the lack of statistical and comparative evidence: the headline result is a single run, and the paper does not establish that the MPC/latent-model machinery adds value over simpler periodic forcing or that the result is repeatable across initial conditions.","major_comments":[{"comment":"The 12.8% drag reduction is computed from a single closed-loop DNS run over 100 convective times. Since the uncontrolled wake at Re=500 is a limit-cycle flow, the finite-time mean Cd depends on the initial phase and on the transient after control is switched on. No error bars, no repeated runs, and no convergence check for the mean are provided. The headline claim therefore rests on a single point estimate. Please add multiple controlled runs from different initial conditions (or randomized phases) and report run-to-run variability, or at least show a time-resolved moving average with a clear window and a stationarity check.","section":"§3.3, Fig. 11; Eq. (13)"},{"comment":"No comparison is made against a simpler open-loop forcing baseline. The paper states that the control action settles into a quasi-periodic pattern at a frequency that counteracts vortex shedding (§3.3). This raises the possibility that the observed drag reduction is caused by periodic actuation at the shedding frequency, not by the model predictive control or the learned latent dynamics. An open-loop harmonic forcing experiment with the same amplitude and frequency (e.g., sinusoidal blowing/suction at f_sh) is needed to isolate the contribution of the MPC framework. This is load-bearing for the paper's central claim that the framework is effective.","section":"§3.3, Fig. 11; §2.4.2"},{"comment":"The paper acknowledges that, as the wake stabilizes, the model 'departs further from its interpolation region.' Since the MPC cost (Eqs. 13–15) is minimized over model-predicted forces, a biased prediction in that extrapolated regime can lead to suboptimal control actions. Fig. 12 reports the force-prediction error along the horizon, but it does not characterize whether the closed-loop trajectory actually moves outside the training-data distribution. The claim that MPC 'operates effectively despite limitations of the underlying model' would be directly supported by quantifying the closed-loop latent-state distribution relative to the training data (e.g., density overlap or distance from the training manifold), or by a robustness test under model mismatch or perturbed initial conditions. This is a specific, addressable gap.","section":"§3.3, Fig. 12 and surrounding text"}],"minor_comments":[{"comment":"The term labeled 'Mean drag increment' is simply the mean predicted Cd minus the constant Cd,ref. Since Cd,ref is a constant, minimizing this term is equivalent to minimizing the mean Cd; the label is somewhat misleading and could be clarified.","section":"Eq. (13)"},{"comment":"The number of optimization iterations is given as 'typically 5' with a learning rate of 10^-3. Please state the exact value used in the reported closed-loop run and whether the result is sensitive to this choice.","section":"§2.4.2"},{"comment":"The colored prediction points in Fig. 11 are visually prominent but the caption does not explain the color mapping or how the 'predicted' values relate to the 'actual' traces. A more explicit caption or a separate legend would improve readability.","section":"Fig. 11"},{"comment":"Reference [41] is malformed: the author list appears as 'K. D. B. J. Adam, et al.' This should be corrected to Kingma, D.P. & Ba, J. (ICLR 2015).","section":"References"},{"comment":"The horizontal axis is labeled 'Number of sensors (Log scale)' but the base of the log scale is not specified. Using explicit tick labels (1, 2, 4, 8, ...) would be clearer.","section":"Fig. 10"},{"comment":"There is a grammatical issue in 'using the Gym-preCICE [33] wrapper for the [34] coupling library.' Also, all software and acronyms (e.g., Gym-preCICE, preCICE) should be defined at first use.","section":"§2.1"},{"comment":"The data availability statement says materials 'will be made openly available in public repositories upon publication.' Please add repository URLs or a review-access link if possible, and specify the exact datasets/code artifacts that will be released.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. My take: sensible, clearly-written integration of offline-trained LSTM latent dynamics, SHAP sensor selection, and gradient-based MPC on surface-pressure sensors. The components are not new, but the end-to-end combination and the demonstration with four base sensors are a real step toward deployable AFC. The drag reduction is measured in the DNS plant, so no definitional circularity. The SHAP rankings are physically sensible, the latent space is interpretable, and the force decoder does well on a chirp test (R2 about 0.94 for Cd and 0.96 for Cl). Credit where it's due.\n\nThe soft spots are in the evidence base, not the methodology. The 12.8% figure rests on a single closed-loop trajectory. No repeated runs, no uncertainty quantification, and 100 tc may not be long enough to fix mean Cd to three digits. The authors honestly note in §3.3 that as the wake stabilizes the model 'departs further from its interpolation region' (Fig. 11). Since MPC minimizes over model-predicted Cd, that is exactly where the optimized actions could be suboptimal. That doesn't kill the result, but it means the headline number is not yet a stable property of the framework.\n\nThe bigger omission is baselines. No comparison with simple open-loop periodic forcing at the shedding frequency, no 90-sensor MPC comparison, and the sensor ablation is evaluated only through open-loop force estimation, not closed-loop drag. Without those, we cannot tell how much of the 12.8% comes from the latent-model MPC machinery versus actuation near the natural shedding frequency. The reported MPC weights are empirically tuned (w_smooth=8.0), which is a free-parameter concern, but at least the settings are transparent. The 2D Re=500 setup is far from truck-scale turbulence, so significance is subfield-level; the authors are appropriately cautious. Code and data are promised, not yet available; for reproducibility, that should be fixed.\n\nThe stress-test concern holds up: in this case the single-run risk and missing baseline are real and not rescued by the paper's internal consistency. It is not a fatal flaw, but the quantitative claim needs repeated runs and a periodic-forcing baseline before it can be called robust.\n\nThe paper deserves a serious referee. I'd send it to review and ask for those additions plus code/data release. If I worked on data-driven AFC, I'd cite it. I'd bring it to a reading group. It is not a desk reject.","headline":"A sensible, clearly-written integration of known pieces that gives a plausible 12.8% drag reduction in a 2D DNS, but the headline number rests on one run with no simple forcing baseline; worth refereeing, not yet a robust result.","tokens_in":16756,"tokens_out":3612,"would_cite":true,"duration_ms":31216,"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":"An offline-trained latent MPC cuts truck-wake drag by 12.8%","keywords":["active flow control","model predictive control","LSTM","latent dynamics","sensor selection","SHAP","drag reduction","truck wake"],"falsifier":"Repeat the identical closed-loop experiment at a different Reynolds number (for instance, Re=1500) or with a perturbed control path, and measure the drag reduction; if the latent model's predictions diverge once the wake is stabilised, the reported gain is tied to the single run rather than to the framework.","tokens_in":15764,"feed_emoji":"🚛","tokens_out":4396,"duration_ms":35974,"temperature":0.7,"pith_summary":"This paper claims that a complete closed-loop active flow control system can be assembled entirely from offline data, using only a handful of non-intrusive surface pressure sensors. In two-dimensional DNS of a simplified truck wake at Reynolds number 500, the resulting controller—built from an LSTM encoder, a latent dynamics model, and model predictive control—reduces mean drag by 12.8% and raises base pressure by more than 35%. The authors' central claim is that the wake state can be inferred from surface pressure history, propagated through a low-dimensional latent space, and optimised by cheap gradient-based MPC, with no online learning. If correct, this offers a practical path from demonstrations to deployable real-time control on road vehicles.","feed_headline":"Four pressure sensors plus latent MPC cut truck wake drag by 12.8%","feed_subtitle":"An LSTM latent model and gradient-based MPC run closed-loop from four surface probes alone.","key_machinery":"The central object is the differentiable latent dynamics chain: an LSTM encoder (state estimator), a residual MLP that advances the latent state under a control action, and a force decoder. Because the entire chain supports automatic differentiation, the model predictive controller's cost—mean drag, drag fluctuation, lift magnitude, and control-signal smoothness—can be minimised by exact gradient backpropagation over a 25-step horizon, making the optimisation cheap enough for receding-horizon real-time use. The latent space is regularised to be decorrelated and force-informative, and it develops a V-shaped structure whose arms correspond to opposite signs of lift, reflecting the vortex-shedd","core_discovery":"The paper demonstrates the complete workflow end to end. A 90-sensor pressure array records the wake's pressure footprint during randomly modulated open-loop jet actuation; an LSTM-based temporal encoder maps 32 steps of this history into an 8-dimensional latent state; an MLP dynamics model predicts the latent-state increment under the current jet action; and a decoder maps the latent state to drag and lift. After training, a Shapley-value attribution analysis ranks the sensors and a slim encoder is distilled to operate on only four base-located probes. In closed-loop DNS, the model predictive controller reduces mean drag coefficient from 1.051 to 0.916, raises mean base pressure from -0.441","pith_inferences":["A natural next step is to test robustness: adding measurement noise, gusts, or a Reynolds-number change to the DNS would probe whether the 12.8% figure survives conditions the model was not trained on.","The Shapley-based sensor ranking is never benchmarked against an exhaustive search over four-sensor subsets; a directed comparison would bound how close the selected sensors are to the best possible placement.","The eight-dimensional latent space is hypothesised to hold for a 2D wake at Re=500; whether the same encoder-dynamics-decoder architecture transfers to three-dimensional bluff-body wakes is untested."],"forward_implications":["A controller trained entirely on open-loop actuator data can be deployed closed-loop without online learning, and in this test case it reduces drag by 12.8%.","Four surface pressure probes located near the base are sufficient to reconstruct the wake state for control, eliminating the need for intrusive wake probes.","Because the MPC cost function is explicit, the same trained model could be reused to balance drag reduction against energy consumption or actuator wear at deployment time.","The observed mechanism is wake stabilisation: a longer recirculation bubble, weaker base suction, and reduced transverse velocity fluctuations (standard deviation down more than 23%)."],"fun_headline_variants":["Four probes plus LSTM-MPC trim truck drag by 12.8%","Latent MPC uses 4 pressure sensors to cut truck wake drag 12.8%","LSTM-MPC with sparse sensors achieves 12.8% drag reduction on truck","Smart MPC from 4 sensors delivers 12.8% truck wake drag cut"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The 12.8% reduction rests on the trained latent-dynamics model staying accurate while the controller drives the wake away from the data it was trained on—the paper notes the model 'departs further from its interpolation region' as the wake stabilises, and a single simulation run does not quantify that risk.","fun_headline_variants_meta":{"raw":{"variants":["Four probes plus LSTM-MPC trim truck drag by 12.8%","Latent MPC uses 4 pressure sensors to cut truck wake drag 12.8%","LSTM-MPC with sparse sensors achieves 12.8% drag reduction on truck","Smart MPC from 4 sensors delivers 12.8% truck wake drag cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1209,"prompt_tokens":698,"completion_tokens":511,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":442,"tokens_out":511,"duration_ms":4546,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:04:29.891370+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the identical closed-loop experiment at a different Reynolds number (for instance, Re=1500) or with a perturbed control path, and measure the drag reduction; if the latent model's predictions diverge once the wake is stabilised, the reported gain is tied to the single run rather than to the framework.","supporting_citations":[],"review_version":1}