{"id":"21625b47-dde3-4551-9f75-32f8858bc821","arxiv_id":"2505.14515","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A wind-speed-scheduled linear surrogate of a floating wind turbine, built from state derivatives extracted via splines, reproduces controller response about 48 times faster than OpenFAST and preserves design-space trends.","lead":"This paper proposes a low-fidelity model for floating offshore wind turbines that approximates turbine dynamics using a wind-speed-dependent linear model trained on simulation data. It reports about a 48x speedup over the high-fidelity OpenFAST simulator while reasonably capturing controller response and design-space trends.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 48x speedup and low simulation times are not in question, but the load-bearing claim that DFSM 'preserves the shape' of the DEL design space is supported only by visual inspection of contour plots; no quantitative comparison of rankings or optima is provided.","rationale":"The paper has real strengths: the speedup numbers are explicit (25 s vs 20 min per load case, 5.2 h vs 250 h for the DOE), the spline-derivative construction is reproducible in principle, and the authors honestly acknowledge that Mt,y and DEL are underpredicted. I do not dispute those. But the strongest claim ties the method's value to controller optimization, and the only evidence for that is qualitative contour inspection. The underprediction is large enough that a non-uniform error would change the location of the optimum, and the paper neither rules that out nor quantifies it. This is a missing-evidence concern rather than an internal inconsistency: the model could pass the quantitative test, but on the current record the central optimization claim is not fully supported. The reader flagged the qualitative design-space comparison and the underprediction of DEL, so I partially agree; I differ in locating the load-bearing gap at the missing quantitative design-space metric rather than at the LPV structural assumption itself. The verdict remains CONDITIONAL: the central claim is credible but requires the proposed quantitative check to be fully supported.","tokens_in":20136,"tokens_out":9110,"duration_ms":91410,"concrete_test":"Re-run the 25-point DOE of Sec. 6.2 with a fixed random seed and report, for both DFSM 1 and DFSM 2, the Spearman rank correlation between DFSM and OpenFAST DELt over the 25 points and the Euclidean distance between their argmin xc. If ρ < 0.8 or the optimum shifts by more than one grid step in either design variable, the 'preserving the shape of the DEL design space' claim is falsified; otherwise it is confirmed quantitatively.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 6.2 is the decisive use case: the paper claims DFSM reproduces the DELt design space over controller gains xc = [ωPC, ζPC] (Fig. 14). The contour plots show DFSM DELt values of roughly 50-85 MNm against OpenFAST's 111-135 MNm, a ~30-40% underprediction that is not shown to be a uniform dilation. If the error varies with xc, or with the wind-speed weighting in Eq. (14), the design space can look similar while the optimizer's ranking is wrong. The authors provide no rank-correlation, no optimum-location comparison, no gradient/sensitivity check, and no unscaled overlay; the scaled plots in Fig. 15 hide the bias, and the two DFSMs give visibly different design-space shapes from each other. Since DELt is computed from the least-accurate modeled quantity, Mt,y (Sec. 6.1), this is the weakest link between the model and the claimed optimization utility. The LPV structural assumption may explain the underprediction, but even if the LPV form is adequate, the paper's evidence does not establish that the design-space trend generalizes across controller gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a linear-parameter-varying (LPV) derivative function surrogate model (DFSM) for the IEA 15-MW floating offshore wind turbine. State derivatives are extracted from OpenFAST simulation time series using cubic splines, and wind-speed-scheduled state-space matrices are fit by solving an optimization problem with a stability constraint. The DFSM is compared in closed-loop simulation with a subspace-identification model (n4sid) and an LSTM model, then evaluated for two use cases: predicting key time series and performance metrics (DELt, AEP) over multiple wind speeds, and reproducing the DELt design space over blade-pitch controller gains (omegaPC, zetaPC) in a 25-point design-of-experiments study. The authors report that the DFSM offers the best trade-off among the three low-fidelity models and provides roughly a 48x speedup over OpenFAST while preserving the qualitative shape of the DELt design space.","tokens_in":20389,"tokens_out":5504,"duration_ms":57119,"significance":"If the quantitative claims survive scrutiny, the LPV-DFSM is a useful engineering contribution: it is a continuous-time, stable-by-construction surrogate with physically meaningful states, and it can be coupled to a controller for closed-loop simulation. The paper deserves credit for evaluating the DFSM on held-out seeds and, in the DOE study, on controller gains not used in training; that generalization test is genuinely external. The main significance limitation is that the central design-space-preservation claim rests on visual comparison of contour plots and on an uncontrolled comparison of the three modeling approaches, so the claimed utility for controller optimization is not yet established at the level of a journal publication.","major_comments":[{"comment":"The headline comparison of the three modeling approaches is not controlled: the DFSM is trained on simulations at all seven wind speeds (5 seeds each), the n4sid model is trained only at 14 m/s, and the LSTM model is trained at 12, 14, and 16 m/s. Fig. 9 then compares their closed-loop blade-pitch MSE on ten 14 m/s test cases and concludes that the DFSM achieves the best accuracy/simulation-time trade-off. Because training data coverage strongly affects both accuracy and variance, this comparison does not isolate modeling approach; please retrain the n4sid and LSTM baselines on the same 7-wind-speed training set (or at least on the same subset as the DFSM) and repeat the MSE/time comparison, or explicitly reframe the claim as 'best under the given per-model training-data budget.'","section":"Sec. 5.3"},{"comment":"The central optimization-utility claim is that the DFSM 'preserves the shape' of the DELt design space, but this is supported only by visual inspection. Fig. 14 shows DFSM DELt levels of roughly 50-85 MNm against OpenFAST's 111-135 MNm, a 30-40% underprediction, and Fig. 15 removes the bias by normalizing each panel to a different scale, making the shapes look more similar than an unscaled overlay would. The two DFSMs trained from different xc values also produce visibly different normalized contours. Since DELt in Eq. (14) is computed from Mt,y, the quantity the DFSM predicts least accurately (Fig. 13c), the paper needs a quantitative evaluation of design-space fidelity: report a rank correlation (e.g., Spearman) between DFSM and OpenFAST DELt values over the 25 DOE points, the location and value of the predicted optimum xc, and an unscaled error map; otherwise the conclusion that the DFSM can be used in controller optimization is not established.","section":"Sec. 6.2, Figs. 14-15"},{"comment":"In the closed-loop validation, the only plotted comparison between low-fidelity models and OpenFAST is the blade-pitch signal beta. However, the ROSCO controller uses omega_g and x_ddot_t as feedback variables (Sec. 2.2), so errors in those variables are the direct cause of any beta error; the sentence in Sec. 6.1 that 'the efficacy of the DFSM in predicting omega_g and x_ddot_t can be inferred from Fig. 11a' is not a substitute for a direct comparison. Please add time-series or error metrics for omega_g and x_ddot_t in the closed-loop validation (and in Sec. 6.1), since these are the quantities the plant model must provide for correct control action.","section":"Sec. 5.3"}],"minor_comments":[{"comment":"The arXiv title promises 'optimal control and control co-design', but the manuscript does not perform a co-design study and Sec. 7 explicitly lists nested control co-design as future work; please align the title and abstract with the actual scope.","section":"Title/Abstract"},{"comment":"The LSTM model is stated to predict y_k = f_LSTM(u_k) with a 0.5 s timestep while the other models use dt = 0.01 s; please specify whether the LSTM was trained on downsampled data and how outputs are compared at common time instants.","section":"Sec. 5.2"},{"comment":"The scatter plot of MSE versus simulation time would be easier to read on a logarithmic time axis, since the DFSM and LSTM points at 25 s and 70 s are compressed against the n4sid points at 2.5 s.","section":"Sec. 5.3, Fig. 9"},{"comment":"In the definition of DELt, the dependence of DEL(wbar) on the seed and on the controller gains xc is not made explicit; please clarify the notation so that the weighted integral is unambiguous.","section":"Sec. 6.1, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is defensible and the DOE generalization test is a genuine strength, but the two load-bearing gaps (uncontrolled training-data comparison and lack of quantitative design-space fidelity metrics) are fixable in revision. I recommend major revision and would encourage the authors to provide the data and model artifacts, or at least the quantitative DOE metrics, to make the journal version self-contained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this is a decent engineering paper rather than a breakthrough, and I think it deserves a proper review. The new piece is the LPV-based derivative function surrogate: they extract state derivatives from OpenFAST/ROSCO time series by cubic spline, then fit wind-speed-scheduled linear state-space matrices with a stability constraint, using the 14 m/s optimum as a warm start for neighboring wind speeds. That last trick cuts construction time from 460 s to ~2.6 s for each extra wind speed, which is a nice practical point. The comparison against n4sid and LSTM is reasonable in spirit and the paper is honest about the main weakness: the DFSM underpredicts tower-base DEL by 30-40% and the authors say so.\n\nThe biggest soft spot is the evidence for the paper's load-bearing claim that the DFSM preserves the shape of the DEL design space in the controller gain coordinates (omega_PC, zeta_PC). That rests on visual inspection of contour plots. The scaled plots in Fig. 15 hide the actual bias, and the two DFSMs trained from different operating points give different design-space shapes. No rank correlation, no comparison of the optimum location, no gradient check. Since DELt is built from Mt,y, the quantity the model fits least well, the trend claim is weaker than the paper implies. That is fixable: a few numbers would do.\n\nThere are also two smaller issues. The baseline comparison is not apples-to-apples: the DFSM sees data from all seven wind speeds, n4sid sees only 14 m/s, and LSTM sees three wind speeds, so the 'better balance' conclusion is partly a statement about training data. And the closed-loop validation reports only blade pitch; it does not directly validate the feedback variables (omega_g and tower-top acceleration) that the controller uses, even though the paper says those are captured well. Those are addressable, not fatal.\n\nThe LPV structural assumption—six states, wind speed as the only scheduling variable, states drawn from measured outputs—is a real limitation and the paper acknowledges it. For early-stage controller tuning, it is probably okay, but the generalization across gains should be tested more rigorously.\n\nWho is this for? Control engineers working on floating wind turbine co-design who want a fast closed-loop plant model. The speedup claim (about 48x) is solid and not in dispute. I would cite it if I were in that area, and I would bring it to a reading group, but I would want the authors to add the quantitative design-space comparison and release the data/code.\n\nMy recommendation: send to peer review, with the expectation of major-ish revisions focused on the DOE evidence and the fairness of the baseline comparison.","headline":"A useful incremental surrogate method for FOWT controller optimization, with a credible 48x speedup but a design-space trend claim that needs quantitative support.","tokens_in":20916,"tokens_out":2561,"would_cite":true,"duration_ms":23111,"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":"A wind-speed-scheduled linear surrogate of floating offshore wind turbine dynamics, trained on simulated state-derivative data, reproduces the controller design-space shape at nearly 48x lower simulation cost and beats subspace…","keywords":["derivative function surrogate model","linear parameter-varying systems","floating offshore wind turbines","data-driven low-fidelity modeling","system identification","long short-term memory networks","damage equivalent load","controller optimization"],"falsifier":"Train the same DFSM on a floating turbine with a different platform or controller architecture and compare the damage-equivalent-load contour over controller gains with the high-fidelity simulator; if the contour shape changes or the predicted optimal gain moves outside the contour spacing, the LPV structure is insufficient. A sharper test is to simulate the DFSM-chosen optimal controller in the high-fidelity model and check whether its damage-equivalent load is close to the true optimum across several wind seeds.","tokens_in":19924,"feed_emoji":"🌊","tokens_out":11542,"duration_ms":103424,"temperature":0.7,"pith_summary":"This paper proposes building a derivative function surrogate model (DFSM) for a floating offshore wind turbine by fitting a wind-speed-scheduled linear state-space model to state-derivative data extracted from high-fidelity simulations. The authors argue that this linear parameter-varying DFSM balances simulation time and prediction accuracy better than a subspace state-space identification baseline (n4sid) or an LSTM network, reducing a roughly 20-minute closed-loop high-fidelity simulation to about 25 seconds. They show that the DFSM reproduces the shape and trends of the tower-base damage-equivalent-load design space over pitch-controller gains, even though it underpredicts the absolute damage value. If correct, the approach makes multi-hundred-evaluation controller optimization studies for floating offshore wind turbines feasible on a personal computer rather than a computing cluster.","feed_headline":"Surrogate model speeds floating wind turbine optimization 48-fold","feed_subtitle":"Scheduled by wind speed, the linear model runs in 25 seconds and keeps the DEL design-space shape of the full simulator.","key_machinery":"The load-bearing machinery is the derivative function surrogate model built through a spline-derived identification pipeline. Simulated state trajectories are approximated by cubic splines, and the exact polynomial derivatives give approximate state-derivative time series. For each mean wind speed, an optimization problem fits the state matrices $(A(w), B(w))$ to minimize the derivative error subject to the constraint that all eigenvalues of $A(w)$ have negative real part, and a second least-squares problem fits $(C(w), D(w))$ to the measured outputs. Because wind-speed trajectories overlap, the converged solution for one mean wind speed seeds the next, so the identification cost drops from 460 seconds for the first wind speed to about 2.6 seconds for each neighbor. Interpolating these matrices across wind speed yields the continuous LPV plant that is then coupled with the reference open-source controller for closed-loop simulation.","core_discovery":"The central claim is that a continuous-time linear parameter-varying model, with six physically meaningful states (platform pitch, tower-top displacement, generator speed, and their first derivatives) and wind speed as the sole scheduling variable, can serve as an accurate low-fidelity plant for control studies. The system matrices $A(w)$, $B(w)$, $C(w)$, $D(w)$ are identified by minimizing the error between spline-extracted state derivatives and model predictions, with a stability constraint on the eigenvalues of $A(w)$, followed by a least-squares fit of the output map. In closed-loop tests the surrogate tracks blade pitch and generator power closely, captures the mean of the tower-base moment but underpredicts its range, and produces a damage-equivalent-load contour over controller gains whose shape matches the high-fidelity simulator even when the DFSM was trained at a different controller gain. The paper's quantitative result is a nearly 48-fold speedup, about 25 seconds per simulation instead of roughly 20 minutes, with lower response variance than n4sid and lower simulation time than the LSTM baseline.","pith_inferences":["Beyond the paper, the same spline-derivative-plus-LPV pipeline should transfer to other expensive dynamic simulators, such as wave energy converters or marine hydrokinetic turbines, whenever the scheduling variable is measurable and the key states appear among the outputs.","The known shortfall—underprediction of the tower-base moment range—suggests a testable extension: keep the linear state equation but add a small static nonlinear correction to the output map; if the DEL contour improves without losing the speedup, the output map, not the derivative identification, was the limiting factor.","A stronger validation the paper leaves implicit is to run the DFSM-selected optimal controller in the high-fidelity simulator and compare its DEL to the true optimum; if the gap stays small across seeds, the surrogate could replace full-fidelity DOE in early design stages.","The authors note the DFSM was trained at one controller gain point; a plausible generalization is that design-space fidelity degrades gradually as the training gain moves away from the optimum, which could be checked by training at several gains and comparing contour shapes."],"forward_implications":["A 25-point design-of-experiments scan of pitch-controller gains drops from roughly 250 CPU-hours with the high-fidelity simulator to about 5 CPU-hours with the DFSM, turning a cluster job into a workstation task.","The DFSM supplies closed-loop time series of blade pitch, generator power, and tower-base moment, so performance metrics such as damage equivalent load and annual energy production can be post-processed from its outputs.","The surrogate predicts responses for controller gains it was never trained on and preserves the shape of the DEL design space, making it usable for early-stage controller optimization before high-fidelity verification.","The DFSM is a continuous-time model with physically meaningful states, unlike the n4sid baseline whose states are abstract and whose required model order changes with wind speed, which blocks a direct LPV extension.","Training the DFSM takes about 509 seconds on top of roughly 11.6 CPU-hours of simulation data, comparable to LSTM training time but with much faster evaluation."],"supporting_citations":[{"why":"Introduces derivative function surrogate models, the modeling class the paper extends with an LPV identification scheme.","marker":"[1]"},{"why":"Provides the trajectory-based sampling and sequentially refined metamodel strategy for DFSMs that motivates the time-series data setup.","marker":"[2]"},{"why":"Earlier LPV control co-design study for the same turbine class; supplies the spline-based state-derivative extraction the paper reuses and validates.","marker":"[9]"},{"why":"Defines the reference open-source controller whose pitch-loop gains serve as the design variables and whose closed-loop structure is used in validation.","marker":"[11]"},{"why":"Supplies the subspace identification theory underlying the n4sid baseline model.","marker":"[30]"},{"why":"SSARX subspace identification method that handles closed-loop data, used to construct the n4sid baseline.","marker":"[34]"},{"why":"Definition of the IEA 15-MW reference wind turbine used as the system under study.","marker":"[43]"},{"why":"Definition of the semisubmersible platform model coupled to the 15-MW turbine.","marker":"[44]"},{"why":"Provides the rainflow-counting and Palmgren-Miner procedure used to compute the damage equivalent load metric.","marker":"[50]"}],"fun_headline_variants":["LPV surrogate cuts floating wind turbine optimization time 48x","48x faster optimization for floating wind turbines with LPV model","LPV surrogate model: 48x speedup for FOWT design optimization","Wind-speed-scheduled LPV model makes FOWT optimization 48x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The turbine's dynamics are assumed to be fully represented by six measured states evolving under a linear model whose matrices depend only on the instantaneous wind speed, with no unmeasured internal states.","fun_headline_variants_meta":{"raw":{"variants":["LPV surrogate cuts floating wind turbine optimization time 48x","48x faster optimization for floating wind turbines with LPV model","LPV surrogate model: 48x speedup for FOWT design optimization","Wind-speed-scheduled LPV model makes FOWT optimization 48x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2782,"prompt_tokens":947,"completion_tokens":1835,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":1756}},"tokens_in":563,"tokens_out":1835,"duration_ms":11434,"temperature":1.0,"reasoning_tokens":1756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:32:23.722763+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same DFSM on a floating turbine with a different platform or controller architecture and compare the damage-equivalent-load contour over controller gains with the high-fidelity simulator; if the contour shape changes or the predicted optimal gain moves outside the contour spacing, the LPV structure is insufficient. A sharper test is to simulate the DFSM-chosen optimal controller in the high-fidelity model and check whether its damage-equivalent load is close to the true optimum across several wind seeds.","supporting_citations":[{"cited_title":"A trajectory-based sampling strategy for sequentially refined metamodel management of metamodel-based dynamic optimization in mechatronics","cited_arxiv_id":null,"evidence_quote":"Provides the trajectory-based sampling and sequentially refined metamodel strategy for DFSMs that motivates the time-series data setup."},{"cited_title":"A reference open-source controller for fixed and floating offshore wind turbines","cited_arxiv_id":null,"evidence_quote":"Defines the reference open-source controller whose pitch-loop gains serve as the design variables and whose closed-loop structure is used in validation."},{"cited_title":"V ., and Moor, B","cited_arxiv_id":null,"evidence_quote":"Supplies the subspace identification theory underlying the n4sid baseline model."},{"cited_title":"A new subspace identification method for open and closed loop data","cited_arxiv_id":null,"evidence_quote":"SSARX subspace identification method that handles closed-loop data, used to construct the n4sid baseline."},{"cited_title":"Definition of the UMaine V olturnUS-S ref- erence platform developed for the IEA wind 15-megawatt o ffshore reference wind turbine","cited_arxiv_id":null,"evidence_quote":"Definition of the semisubmersible platform model coupled to the 15-MW turbine."},{"cited_title":"Determining equivalent dam- age loading for full-scale wind turbine blade fatigue tests","cited_arxiv_id":null,"evidence_quote":"Provides the rainflow-counting and Palmgren-Miner procedure used to compute the damage equivalent load metric."}],"review_version":1}