{"id":"e99617c9-bfe0-4695-86ed-1fda79319fca","arxiv_id":"2502.09531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DeePC, a purely data-driven predictive controller, achieves angle tracking and vibration suppression for a simulated flexible spacecraft without using any model of the flexible dynamics.","lead":"A data-driven predictive controller is applied to a simulated flexible spacecraft, where a torque at the hub must rotate a long flexible appendage to a target angle while damping its vibrations. The authors report that this model-free controller matches or beats a Lyapunov-based boundary controller in cost, with a faster settling time when the model parameters are wrong.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fundamental lemma's persistent-excitation precondition for the offline PD-collected data is never verified, and the SVD truncation can mask rank deficiency; without this verification the DeePC surrogate in (12) lacks theoretical justification, so the simulation evidence does not establish the…","rationale":"The reader's weakest_assumption identifies the unverified persistence of excitation and the identification of the FE discretization with the true plant. My stress-test concurs and sharpens the concern: not only is the condition unverified, but the specific data-collection protocol (PD tracking of step references) makes the condition implausible for a high-order FE system, and the SVD truncation in Section III.B actively hides the rank deficiency by projecting onto the first r singular vectors. This is load-bearing because the entire DeePC construction in (12) inherits its validity from Lemma 1; if the Hankel matrix does not span all trajectories, the non-parametric model is not a surrogate for the system, and the simulation results, while real, are not evidence for the claimed framework. The output-definition ambiguity and the undefined cost function in Table I are also serious and would hamper reproduction, but they concern the comparison and performance claims; the PE failure attacks the theoretical foundation of the method itself. A concrete rank check and a re-run with a provably PE input would settle the question: if the method is robust to the data-collection excitation, the concern is mitigated; if performance degrades, the original success was an artifact of the particular PD-generated data. The reader's CONDITIONAL verdict remains appropriate: the paper is a plausible feasibility demonstration, but the central claim is conditional on verifying the lemma's precondition and clarifying the protocols, not on a fatal flaw. I therefore recommend no change to the reader's verdict.","tokens_in":19749,"tokens_out":7345,"duration_ms":77363,"concrete_test":"Extract the offline input sequence u^d from Section IV.A; set L = Tini + N = 40 and compute the input Hankel matrix H_L(u^d) and the combined Hankel matrix in (8). Compute the numerical rank via singular values (e.g., threshold at machine precision times the largest singular value) and compare with the required row rank m(n+L) = n+40 for m=1, where n is the state dimension of the FE model (≈86). If the rank is below this threshold, the fundamental lemma's condition fails. Then, as a control experiment, generate a new offline dataset using a provably persistently exciting input (e.g., a pseudorandom binary sequence or a sum of sinusoids with distinct frequencies spanning the system bandwidth), rebuild the Hankel matrix with the same Tini, N, Q, R, and rerun the nominal scenario.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Willems' fundamental lemma, which requires the offline input sequence to be persistently exciting of order n+L, where n is the state dimension of the data-generating LTI system. Section IV.A collects data under a PD controller tracking 'several desired angles,' and Section III.A asserts the data is 'sufficiently rich' without ever computing the full-row-rank condition in Definition 1. For the FE-discretized plant (Appendix I) with 20 elements, the state dimension n is roughly 86, and L = Tini + N = 40, so the input Hankel matrix must have full row rank at least n+L = 126. A PD tracking response to step references is a damped transient that typically excites only a few frequencies and is very unlikely to satisfy this condition. The SVD reduction in (15)-(16) truncates the Hankel matrix to rank r = 80, discarding small singular values, so even if the original matrix is rank deficient, the reduced problem remains feasible; the slack variable σy and regularization λg‖g‖² further mask any violation of the trajectory-spanning property. Consequently, the equality constraint in (12) may not represent all input-output trajectories, and the optimization solved in simulation is not guaranteed to be the DeePC controller justified by the fundamental lemma. Without verifying persistence of excitation, the reported tracking and vibration suppression, while plausible, are not explained by the paper's theoretical framework, and the conclusion that DeePC 'demonstrates its capability' model-free is unsupported. The same FE discretization is used for data generation and closed-loop evaluation, so the infinite-dimensional PDE-ODE plant of Section II is never actually controlled; this is a secondary but related limitation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-enabled predictive control (DeePC) scheme for a flexible spacecraft modeled as a central rigid hub with a flexible Euler–Bernoulli appendage. It collects offline input-output data under a PD controller, builds a Hankel data matrix, applies an SVD-based dimension reduction, and solves a regularized receding-horizon optimization. The controller is evaluated in a 20-element finite-element simulation and compared with a Lyapunov boundary controller in nominal, model-uncertainty, and process-noise scenarios. The reported results show comparable or better angle tracking and lower cost for DeePC, and the paper concludes that the simulations provide strong evidence for DeePC's validity and effectiveness. The manuscript also contains a detailed Lyapunov stability proof and a finite-element discretization appendix.","tokens_in":20077,"tokens_out":7400,"duration_ms":73174,"significance":"The application area is relevant and the FE simulation study is a useful first demonstration of DeePC for flexible spacecraft. The benchmark Lyapunov derivation in Appendix II is unusually detailed and constructive, and the FE model in Appendix I is coherent. However, the claims as stated outpace the evidence: the theoretical justification via the fundamental lemma is not verified for the collected data, the 'infinite number of vibration modes' claim is not supported by a finite-element study, and the reported superiority rests on an undefined cost function and hand-tuned hyperparameters. With these issues addressed, the empirical contribution would be a reasonable feasibility study for data-driven boundary control of flexible structures.","major_comments":[{"comment":"The persistent-excitation requirement is asserted but never verified. For the 20-element FE model of Appendix I, the state dimension n is on the order of 82, and with Tini=20, N=20, the fundamental lemma would require the scalar input to be persistently exciting of order n+L, where L=40. The offline data are collected under a PD controller tracking step commands; a damped transient response to a few setpoints is not obviously sufficiently rich, and no rank test of the Hankel matrix in (8) is reported. Because the SVD reduction in (15)-(16) with r=80 and the slack variable sigma_y in (12) can preserve feasibility even when the Hankel matrix is rank deficient, the optimization actually solved is not guaranteed to be the DeePC problem justified by Lemma 1. The convergence in Figs. 4-9 is therefore not explained by the paper's theoretical framework; the authors should either verify full row rank of the relevant Hankel matrix or use an offline excitation signal with a documented persistent-excitation property.","section":"Section III.A, Definition 1/Lemma 1 and Section IV.A/IV.C"},{"comment":"The claim that DeePC can 'stabilize an infinite number of vibration modes' is not demonstrated and, as written, is not compatible with the methodology. The offline data and closed-loop evaluation are both obtained from the finite-dimensional ODE system (35) in Appendix I with 20 beam elements, not from the PDE-ODE system (1)-(4). The fundamental lemma is a finite-dimensional LTI statement, so the DeePC surrogate in (12) applies, at best, to the discretized model. This contribution claim should be restricted to the tested finite-dimensional discretization or reworded to avoid implying a result for the infinite-dimensional PDE.","section":"Section I and Section V"},{"comment":"The controlled output is not consistently specified. Section III.A states that 'the system output is the tip-end deflection,' while Algorithm 1 and the simulation section refer to a 'spacecraft angle output sequence' and plot both theta(t) and tip deflection. Since L=40 and the Hankel matrix has 80 rows, the data matrix in (8) uses a single scalar output in addition to the scalar input. Whether the DeePC output is theta(t) or y(L,t) is essential for interpreting the claim that DeePC simultaneously tracks angle and suppresses vibration. Please define the output vector unambiguously and state which measurement enters equations (8) and (12).","section":"Section III.A and Algorithm 1"},{"comment":"The 'cost function' used to compare DeePC and the Lyapunov controller is never defined. Equation (12) defines a per-horizon DeePC objective, but the values reported in Table I (e.g., 2287 versus 2741) are presumably cumulative over the full simulation; without the exact formula and normalization, a reader cannot reproduce Table I or judge the claimed superiority. In addition, Q, R, lambda_g, lambda_y, Tini, N, and the rank cutoff r are selected per scenario with no tuning rule and no sensitivity analysis, so the comparison is not robust evidence of a general advantage. Please add the cost definition and at least one sensitivity study for the main hyperparameters.","section":"Section IV.C and Table I"}],"minor_comments":[{"comment":"'xk in R^m is the state vector' should read xk in R^n; the symbol m is already used for the input dimension.","section":"Section III.A, after Eq. (5)"},{"comment":"The caption contains the typo 'comparision' and should read 'comparison.'","section":"Section IV.C, Fig. 8 caption"},{"comment":"In the first term of the displayed expression, theta(t) appears where theta_t(t) is evidently intended; please correct this so the integration-by-parts step is transparent.","section":"Appendix II, Eq. (54)"},{"comment":"The process-noise column omits settling times; state whether settling time is defined only for the nominal and uncertainty scenarios.","section":"Table I"},{"comment":"The phrase 'sufficiently rich dataset' does not connect to Definition 1; a sentence explicitly stating T, Tini, N, and the state dimension used in the persistent-excitation order would help readers verify the Hankel construction.","section":"Section IV.A"},{"comment":"Line 1 lists the inputs but not their lengths; specify T and the sampling interval so that the Hankel matrix construction is fully reproducible.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper would be materially strengthened by releasing the data-generating simulation scripts, since the central evidence is empirical. The 'first application' novelty claim should also be checked against the broader DeePC literature, although I do not see grounds for a novelty dispute. The persistent-excitation verification and the cost-function definition are the key blockers for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a legitimate first application of DeePC to boundary control of a flexible spacecraft, and the central angle-tracking simulation is credible. Second, the claims as written are noticeably stronger than what the evidence supports, mostly in places that are fixable with extra detail and softened language rather than fatal flaws.\n\nWhat is genuinely new: this is the first DeePC application to a PDE-ODE flexible spacecraft system, and the comparison against a Lyapunov boundary controller under nominal, uncertainty, and process-noise cases in a finite-element simulation is not in the cited literature. The SVD dimension reduction from the authors' earlier work is sensibly imported, and the Lyapunov appendix is thorough—the stability proof is real, not ornamental. The FE model description is good enough to reproduce.\n\nNow the soft spots. The stress-test's persistence-of-excitation concern lands. The paper never verifies the full-row-rank condition behind the fundamental lemma, and the data collection is described only as a PD controller tracking several desired angles. That kind of damped transient is unlikely to excite all modes to the required order, especially with FE state dimension around 86 and L = 40. The SVD truncation to r = 80 plus the slack variable can mask rank deficiency, so the optimization actually solved may not be the DeePC controller the theory justifies. That is the main technical gap. The related limitation—that the same FE discretization generates both the data and the closed-loop evaluation—means the infinite-dimensional PDE is never actually controlled, which undercuts the phrase 'stabilize an infinite number of vibration modes' in the introduction. The controlled output is also defined inconsistently: the text says tip-end deflection, the algorithm uses end-tip angle, and Table I's cost function is never defined. The uncertainty and noise scenarios lack protocol details, and there are no error bars or multiple runs.\n\nThese are all addressable. The core simulation result—DeePC tracking and vibration suppression without a model—is plausible and worth taking seriously. The paper just needs to verify or at least discuss the PE condition (or show the reduced Hankel matrix still spans the relevant trajectories), define the output and cost, report the noise/uncertainty setups, and tone down the infinite-mode claim.\n\nThis paper deserves a serious referee. If it lands in your hands, the revision should focus on the PE verification and a consistent, reproducible simulation protocol. I would bring it to a reading group as a useful example of how DeePC is being pushed into new application domains, not as a definitive proof of model-free superiority.","headline":"A credible but over-claimed first application of DeePC to flexible spacecraft boundary control: the simulation is plausible, but the persistence-of-excitation precondition is unverified and several claims outrun the evidence.","tokens_in":20697,"tokens_out":1404,"would_cite":false,"duration_ms":17384,"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":"This paper claims that data-enabled predictive control (DeePC) can steer a flexible spacecraft's hub angle and suppress beam vibrations using only recorded torque–angle data, with no parametric model, and that in finite-element…","keywords":["data-enabled predictive control","DeePC","flexible spacecraft","boundary control","Willems fundamental lemma","data-driven control","vibration suppression","singular value decomposition"],"falsifier":"Compute the rank of the 80 × 3961 Hankel matrix formed from the Section IV.A data and compare it with full row rank; if the rank falls short, the fundamental-lemma condition is violated. Alternatively, train DeePC on a coarse finite-element mesh and test it on a finer mesh or on the PDE-ODE model directly: if tracking or vibration suppression degrades sharply, the reported success is tied to training and evaluation sharing the same discretization.","tokens_in":19513,"feed_emoji":"🛰️","tokens_out":6283,"duration_ms":59600,"temperature":0.7,"pith_summary":"Flexible spacecraft appendages are hard to control because their PDE-ODE dynamics resist accurate modeling and discretization can cause spillover. This paper tries to establish that the data-enabled predictive control (DeePC) framework—which builds a non-parametric model directly from recorded torque and angle trajectories via Willems' fundamental lemma—can replace the model-based Lyapunov boundary controller for this system. The authors report finite-element simulations in which DeePC tracks the desired hub angle and suppresses beam vibration across nominal, model-uncertainty, and process-noise scenarios, with lower cost than the Lyapunov benchmark and faster settling in the uncertainty case. A dimension-reduction step based on singular value decomposition keeps the optimization manageable. A sympathetic reading: the value is that a controller needing no parameter calibration can match, and in uncertain cases beat, a controller that assumes exact knowledge of the plant.","feed_headline":"Data-only controller steers a flexible spacecraft","feed_subtitle":"DeePC tracks the hub angle and quiets beam vibration without knowing stiffness or inertia, beating the model-based baseline.","key_machinery":"The load-bearing object is the Hankel data matrix (8) built from an offline input–output trajectory, used as a non-parametric stand-in for the state-space model. Willems' fundamental lemma states that if the recorded input is persistently exciting of order n+L, every length-L trajectory of the controllable LTI system is a linear combination of the columns of this matrix; DeePC then solves the predictive-control optimization (12), which selects a coefficient vector g, a future input u, and a future output y while penalizing a slack variable for noise and adding an ℓ2 regularizer on g. The paper compresses the Hankel matrix with an SVD, keeping the range space via \\bar{H}_L = H_L V_1 = W_1 \\Sigma_1, which cuts the optimization dimension from about 3960 to 80 in the experiments. The same structure is what lets the controller claim to work without model identification.","core_discovery":"The central claim is that DeePC, using only input–output data collected while a PD controller drives the spacecraft, can achieve accurate angle tracking and vibration suppression for a flexible spacecraft without prior knowledge of system parameters or derivation of a mathematical model. In the paper's finite-element simulations, the regularized DeePC formulation (12) with the SVD-compressed data matrix (16) produces lower cost than the Lyapunov-based boundary controller in the nominal case (2287 vs. 2741), under model uncertainty (2623 vs. 3870), and under process noise (2290 vs. 2801); its settling time is longer in the nominal case (62.8 vs. 58.6 s) but markedly shorter under uncertainty (64.3 vs. 90.6 s). The authors take these results as strong evidence of the validity and effectiveness of the DeePC approach for flexible spacecraft.","pith_inferences":["Editorial inference: the same pipeline should transfer to other flexible structures—solar panels, booms, manipulators—as long as a stabilizing baseline controller can generate persistently exciting data; the paper only demonstrates the antenna-hub configuration.","Editorial inference: because the fundamental lemma applies to finite-dimensional LTI systems, the paper's evidence does not by itself prove performance on the true infinite-dimensional plant; a natural test is to train on one finite-element mesh and evaluate on a finer mesh or on a different discretization order.","Editorial inference: the SVD turning-point cutoff is chosen by inspection of the singular-value distribution; a sensitivity study over the rank r would reveal how much of the reported robustness depends on that choice.","Editorial inference: online or recursive updating of the data matrix could allow DeePC to track slow parameter drift during a long mission, an extension the paper does not pursue."],"forward_implications":["A controller for a flexible spacecraft can be synthesized from torque and tip-angle data alone, eliminating the need to identify stiffness, inertia, and damping parameters.","DeePC matches or outperforms the Lyapunov boundary controller on the reported cost in all three scenarios, and keeps a fast settling time when the true mass and stiffness are doubled.","The SVD compression reduces the DeePC decision variable from thousands of entries to the rank of the data matrix, making the online optimization feasible for longer horizons.","Because the offline data is generated by a simple PD controller, the method only requires a stabilizing, not optimal, baseline for data collection.","The approach inherits the robustness of regularized DeePC: the λ_g‖g‖² penalty is what the authors credit for maintaining performance under model uncertainty and noise."],"supporting_citations":[{"why":"Willems' fundamental lemma, which guarantees that a persistently exciting recorded trajectory spans all trajectories of a controllable LTI system; this is the theoretical basis for the non-parametric model.","marker":"[17]"},{"why":"Introduces the DeePC formulation and its regularized version with slack variable and ℓ2 penalty, which the paper adapts in (12).","marker":"[18]"},{"why":"Supplies the SVD-based dimension-reduction method used to compress the Hankel matrix in (14)-(16).","marker":"[28]"},{"why":"Presents the boundary-control Lyapunov approach that the paper modifies to build its benchmark controller.","marker":"[30]"},{"why":"Provides the equivalent ODE form (33) of the hub-beam coupled dynamics used to build a symmetric finite-element model.","marker":"[25]"},{"why":"Textbook basis for the Euler-Bernoulli beam finite elements and the Generalized-α time integration used in simulations.","marker":"[31]"}],"fun_headline_variants":["Data-only DeePC outperforms model-based control for flexible spacecraft","Model-free DeePC quiets flexible spacecraft vibration","Beating Lyapunov: data-driven control of flexible spacecraft","DeePC eliminates model for flexible spacecraft control","Data-driven control beats Lyapunov method for flexible spacecraft"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the torque sequence used to collect offline data is persistently exciting enough that the recorded input–output Hankel matrix spans every trajectory of the discretized spacecraft; the paper calls the data 'sufficiently rich' but never verifies the full-row-rank condition of Definition 1.","fun_headline_variants_meta":{"raw":{"variants":["Data-only DeePC outperforms model-based control for flexible spacecraft","Model-free DeePC quiets flexible spacecraft vibration","Beating Lyapunov: data-driven control of flexible spacecraft","DeePC eliminates model for flexible spacecraft control","Data-driven control beats Lyapunov method for flexible spacecraft"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000482,"raw_usage":{"total_tokens":2335,"prompt_tokens":852,"completion_tokens":1483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":1404}},"tokens_in":468,"tokens_out":1483,"duration_ms":10475,"temperature":1.0,"reasoning_tokens":1404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:09:03.828887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the rank of the 80 × 3961 Hankel matrix formed from the Section IV.A data and compare it with full row rank; if the rank falls short, the fundamental-lemma condition is violated. Alternatively, train DeePC on a coarse finite-element mesh and test it on a finer mesh or on the PDE-ODE model directly: if tracking or vibration suppression degrades sharply, the reported success is tied to training and evaluation sharing the same discretization.","supporting_citations":[{"cited_title":"A note on persistency of excitation,","cited_arxiv_id":null,"evidence_quote":"Willems' fundamental lemma, which guarantees that a persistently exciting recorded trajectory spans all trajectories of a controllable LTI system; this is the theoretical basis for the non-parametric model."},{"cited_title":"Data-enabled p redictive control: In the shallows of the deepc,","cited_arxiv_id":null,"evidence_quote":"Introduces the DeePC formulation and its regularized version with slack variable and ℓ2 penalty, which the paper adapts in (12)."},{"cited_title":"Dimension reduc tion for efficient data-enabled predictive control,","cited_arxiv_id":null,"evidence_quote":"Supplies the SVD-based dimension-reduction method used to compress the Hankel matrix in (14)-(16)."},{"cited_title":"Bou ndary con- trol of anti-symmetric vibration of satellite with flexible appendages in planar motion with exponential stability,","cited_arxiv_id":null,"evidence_quote":"Presents the boundary-control Lyapunov approach that the paper modifies to build its benchmark controller."},{"cited_title":"Improving regulation of a sing le-link flexible manipulator with strain feedback,","cited_arxiv_id":null,"evidence_quote":"Provides the equivalent ODE form (33) of the hub-beam coupled dynamics used to build a symmetric finite-element model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Textbook basis for the Euler-Bernoulli beam finite elements and the Generalized-α time integration used in simulations."}],"review_version":1}