{"id":"4731b4eb-ecb4-43c5-8d0b-24a2178618e5","arxiv_id":"2601.00178","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A PINN framework that jointly parameterizes oscillator phases and additive controls can enforce a prescribed synchronization level after a target time, matching the control cost of an analytical frequency-compensation baseline in Kuramoto systems.","lead":"The paper trains physics-informed neural networks to jointly produce state and control trajectories that drive coupled oscillators into synchronization within a user-set time and coherence level, demonstrating this on Kuramoto networks. It is relevant as a flexible, offline trajectory-optimization alternative to explicit feedback designs, especially in non-gradient systems where analytic control baselines do not exist.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Soft penalty in Eq. (10) at finite collocation points does not enforce R(t)>=R* for all t>=t*, so 'precise regulation' is unsupported; a forward-time check is needed.","rationale":"The reader's weakest assumption and my concern coincide: Eq. (10) is a soft, finite-sample penalty, not an enforcement of the continuum constraint T_s(R*) <= t*. This is the most load-bearing issue because it sits at the exact point where the paper's novelty is claimed: direct trajectory-level regulation of synchronization time and coherence level. If the constraint is not actually satisfied in continuous time, the 'precise regulation' result is only approximate, and the quantitative scans of E versus R* and t* lose their interpretation as valid synchronization-target trade-offs. The concern is not an attack on the PINN approach in general; it is a request for the standard verification that an open-loop controller be tested by forward simulation of the true dynamics. The paper does include some forward simulation in the noise-robustness section (Sec. IV D), which is a point in its favor, but the deterministic noiseless case is not explicitly verified. Other weaknesses noted by the reader, such as the metric E(t) being closely related to L_reg and the lack of code/error bars, are real but secondary; they affect the strength of the cost-efficiency comparison and reproducibility, not the core feasibility of the method. A conditional verdict remains appropriate: the idea is plausible and the numerical demonstrations are suggestive, but the central continuous-time guarantee needs either a rigorous bound or a direct empirical verification. If the forward-simulation check fails, the claim of precise regulation would be falsified; if it passes, the paper's central contribution is substantiated for the reported experiments, though still without formal guarantees.","tokens_in":12771,"tokens_out":7433,"duration_ms":88597,"concrete_test":"Reproduce the Sec. IV A baseline (N=10, K=0.05, R*=1, t*=2), extract the trained control signal u_i(t), and integrate Eq. (5) forward with a high-accuracy adaptive solver over t in [0,T]. Use dense output or event detection to compute min_{t>=t*} R(t). If this quantity falls below R* by more than a numerical tolerance, the 'precise regulation' claim is falsified. Additionally, rerun training or evaluation with doubled and tripled collocation-point counts; if the minimum gap shrinks substantially, the concern is specifically about finite collocation resolution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism for realizing T_s(R*) <= t* is the control loss L_control in Eq. (10): max(0, R* - R(t))^2 evaluated at finitely many collocation points. A small or zero value at those points says nothing about the continuum between them. The paper states in Sec. III that this 'enforces the persistence condition R(t) >= R* for all t >= t*', but no bound, Lipschitz estimate, or denseness argument is provided to justify that jump from a finite set of samples to a continuous-time guarantee. The problem is most acute for the baseline R* = 1 (Sec. IV A), because exact full synchronization is a codimension-(N-1) condition on the phase vector; the network can dip slightly below R=1 between collocation times while the penalty is exactly zero at the sampled times. Compounding this, no post-training forward simulation of the deterministic noiseless Eq. (5) under the learned open-loop control is reported, so the headline plots in Figs. 2-7 may reflect the properties of the network's internal trajectory rather than the actual controlled system. Without such a check, the central claim of simultaneous, precise control of synchronization time and coherence level is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed neural network (PINN) framework for synchronization control of networked oscillators. State and control trajectories are jointly represented by neural networks with time as the only input; the governing dynamics are imposed through residual losses at collocation points, and synchronization targets (threshold R* and target time t*) are encoded via a soft penalty on the order parameter after t*. The authors demonstrate the method on Kuramoto oscillators, reporting smooth controls, reduced transient peaks relative to phase feedback, and cumulative costs comparable to frequency-compensation baselines. They also report extensions to frustrated Kuramoto–Sakaguchi dynamics, Stuart–Landau oscillators, larger networks, and noisy execution. The central claim is that this provides a trajectory-level way to prescribe synchronization time and coherence without assuming an explicit feedback law or solving a strict optimal control problem.","tokens_in":13110,"tokens_out":5616,"duration_ms":59686,"significance":"If the results are reproducible and the continuous-time constraints are actually met, the framework offers a useful alternative to feedback-law design for synchronization regulation, especially in non-gradient systems where analytical baselines are unavailable. The idea of directly penalizing order-parameter persistence is natural, and the parameter studies in Sec. IV B (frequency heterogeneity, coupling strength, target level, target time) are informative. The manuscript's strengths include the breadth of numerical demonstrations (gradient, frustrated, amplitude dynamics) and the honest qualification of the frequency-compensation benchmark as not a universal optimality reference. However, the paper currently lacks the implementation detail and validation needed to support its central claims: no architecture, collocation scheme, loss weights, optimizer, or code are provided, and no post-training forward simulation of the controlled ODE is reported. No machine-checked proofs or reproducible artifacts are included, which limits verification of the numerical conclusions.","major_comments":[{"comment":"The loss L_control penalizes max(0,R* - R(t))^2 only at a finite set of collocation points. The statement in the same section that this 'enforces the persistence condition R(t) >= R* for all t >= t*' is not supported: no Lipschitz/denseness argument is given, and no post-training integration of Eq. (5) with the learned open-loop u(t) is reported. For R*=1, exact phase alignment is a codimension-(N-1) condition, so small inter-collocation dips are plausible. Please report a dense forward simulation of the controlled dynamics after training and quantify the maximum violation of R(t) >= R* on [t*, T].","section":"Sec. III, Eq. (10)"},{"comment":"L_reg = <N^{-1} sum_i u_i(t)^2>_t is the time-averaged control cost, and E = integral_0^T P(t) dt is T times that average (for uniform collocation). The paper's claim that L_reg 'is not interpreted as a physical or performance metric' is contradicted by its use of E as the headline comparison: the PINN is explicitly trained to minimize L_reg, so the reported E is optimized by construction. The comparison with baselines would be meaningful only if E were evaluated on a held-out forward simulation and if the loss weights lambda_control/lambda_reg were reported and varied; otherwise the 'competitive cumulative cost' claim is circular.","section":"Sec. III Eq. (12) and Sec. IV C Eq. (14)"},{"comment":"In the noise-robustness experiment, the control u_i(t) is fixed after training, and P(t) in Eq. (13) depends only on u_i(t), not on the noisy state. Therefore E = integral P(t) dt in Eq. (14) is independent of sigma. Figure 8(b) shows E increasing with sigma, which is impossible under the stated protocol. Please clarify whether the controls were re-optimized for each noise level or redefine the cost to include state-dependent effort; as written, the robustness claim and the cost trend are internally inconsistent.","section":"Sec. IV D, Eq. (19)"},{"comment":"The numerical section does not specify the network architecture (layers/widths/activations), the collocation point set (number/distribution), the loss weights lambda_dyn, lambda_ic, lambda_control, lambda_reg in Eq. (4), the shaping function h(t), the initial control u(0), or the optimizer/hyperparameters. Baseline gains are selected 'by scanning' with no quantitative criterion. Without these details — and without code or a supplementary appendix — the central numerical results cannot be reproduced or independently evaluated. Please provide complete implementation details and, ideally, a reproducibility package.","section":"Sec. II and Sec. IV"}],"minor_comments":[{"comment":"Eq. (14) defines E as a scalar integral, but the text and figures repeatedly write E(t) (e.g., Figs. 4–8). Please use consistent notation; E is not a function of time as defined.","section":"Sec. IV B, Eq. (14)"},{"comment":"Typo: 'reported in the in Supplementary Sec. S3.' Also, the Supplementary Material is referenced (S1–S3) but not included in the submission; the paper should be self-contained or the supplementary sections should be provided.","section":"Sec. IV (introductory paragraph)"},{"comment":"The definition of T_s(R*) requires R(t') >= R* for all t' >= t, but the finite terminal time T is not discussed in relation to this definition. If control is only applied on [0,T], the persistence condition should be stated on [t*,T] or the horizon should be extended; otherwise the condition is vacuous beyond T.","section":"Sec. III, Eq. (7)"},{"comment":"The baseline gain selection criterion is vague ('minimal value that reliably yields a synchronized state with R approximately 1'). Please specify the threshold, duration over which R must remain above it, and the number of initial-condition/realization trials used for the scan.","section":"Sec. IV C"}],"recommendation":"major_revision","confidential_remarks":"To the editor: I see a potentially useful framework, but the manuscript as submitted is not acceptable because two load-bearing claims — continuous-time persistence and the noise-dependence of E — are not supported by the stated methods, and the L_reg/E relationship makes the headline cost comparison circular. These issues are fixable with additional experiments (forward simulation checks, proper noise protocol, full implementation disclosure), so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a sensible core—using a PINN to jointly learn trajectories and controls with a soft constraint on the order parameter after a target time—but two things undercut the headline claims. First, the persistence constraint is only penalized at collocation points, so reaching R*=1 at those points says nothing about the continuum between them; no forward simulation of the controlled system is reported to check. Second, the integrated control cost E(t) is a time-integral of the same quadratic control penalty that the network is trained to minimize (L_reg), so the finding that PINN control is cost-competitive is partly by construction.\n\nWhat's new: the trajectory-level enforcement of a persistent order-parameter inequality for synchronization control is a legitimate extension of PINN-control work (e.g., Mowlavi-Mahadevan) and oscillator control (Zlotnik et al.). The paper is clearly written, the Kuramoto and Kuramoto-Sakaguchi demonstrations are relevant, and the comparison to frequency compensation—which is a genuine analytic benchmark—is thoughtful. The extension to non-gradient, frustrated dynamics is a useful data point.\n\nWhere it's soft: (1) No code, architecture, collocation counts, or hyperparameters; a methods paper needs those. (2) Error bars are shown in Figs. 4-5 but not for the headline benchmark figures, so the \"closely approaches\" claim is unquantified. (3) The paper's own statement that L_reg is \"not a performance metric\" is contradicted by the definition of E(t). That's not fatal—minimizing a cost and then measuring it is fine if the baselines don't optimize it—but the language about \"discovering\" efficiency overstates the finding. (4) The R*=1 target is especially fragile under a soft penalty; exact full synchronization is a codimension-(N-1) condition, so \"precise regulation\" is not supported without a forward-time check.\n\nThe good news is these are all addressable. The central idea survives the critique; it just needs honest reframing and a simple post-training simulation of Eq. (5) under the learned open-loop control.\n\nWho it's for: researchers working on PINN-based control or synchronization control. It deserves a serious referee, but a careful one who will push for reproducibility. I would not cite it in its current form.\n\nRecommendation: send to peer review, but expect major revision.","headline":"A sensible PINN trajectory-optimization idea for synchronization control, undermined by a soft constraint that isn't validated in forward time and a cost metric that is essentially the training regularizer.","tokens_in":13552,"tokens_out":2427,"would_cite":false,"duration_ms":26244,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34C15","68T07","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed neural network can control both the time at which oscillators synchronize and the level of coherence achieved, without prescribing a feedback law.","keywords":["synchronization control","physics-informed neural networks","Kuramoto oscillators","order parameter","trajectory-level control","Kuramoto–Sakaguchi","open-loop control","control cost"],"falsifier":"Take the trained control law, integrate the deterministic Kuramoto dynamics with a fine ODE solver, and check directly whether R(t) ≥ R* for every t in [t*, T]. If any dip below R* exists, the persistence constraint is violated and the claimed precise regulation fails; a specific probe would test R* = 1, where even tiny phase deviations drop R below 1.","tokens_in":12582,"feed_emoji":"🌀","tokens_out":3139,"duration_ms":32855,"temperature":0.7,"pith_summary":"This paper proposes a physics-informed neural network (PINN) approach to synchronization control. Instead of designing a feedback law first and then checking the achieved coherence, the network directly learns open-loop control signals that force the system's order parameter R(t) to stay above a chosen threshold R* for all times after a target time t*. The central claim is that this joint trajectory-control optimization can regulate both synchronization time and coherence level in one training problem. The author shows the learned controls are smooth, avoid the large early-time transients of phase feedback, and closely match the cost of an analytically optimal frequency-compensation strategy in gradient systems. The same framework remains effective in non-gradient frustrated dynamics where analytic baselines break down.","feed_headline":"Neural network sets when and how strongly oscillators sync","feed_subtitle":"Physics-informed training learns smooth open-loop control that matches analytic baselines in gradient and frustrated dynamics.","key_machinery":"The persistence constraint on the Kuramoto order parameter, L_control = <1_{t>=t*} max(0,R*-R(t))^2>, is the mechanism that encodes the synchronization objective. It is combined with a physics-informed residual that enforces the networked phase dynamics and a shaping function h(t) that exactly satisfies initial conditions. Together they turn synchronization regulation into a differentiable trajectory optimization problem with no prescribed control structure.","core_discovery":"The paper's central claim is that macroscopic synchronization objectives can be imposed as persistence conditions on the order parameter at the trajectory level, enabling simultaneous control of synchronization time and coherence level without assuming any explicit feedback law or solving a strict optimal control problem. In the Kuramoto testbed, the physics-informed network jointly parameterizes phases and control inputs, penalizes the dynamics residual, and applies the soft constraint L_control = <1_{t>=t*} max(0,R*-R(t))^2>. The learned control produces smooth synchronization with reduced transient effort, and its cumulative cost approaches the analytic frequency-compensation baseline; in","pith_inferences":["The soft penalty formulation suggests a natural test: how much does the learned control degrade if R* is very close to 1, where exact phase alignment requires increasingly precise control? The sharp rise in cost reported at R*→1 hints the method may need a hard constraint or higher collocation density there.","Because the control is open-loop and trained offline, online adaptation could be added by re-training on the observed state; the paper's robustness results suggest a natural extension to model-predictive control with PINN warm-starts.","One could extend the same persistence idea to other collective observables (e.g., chimera measure, phase velocity variance), turning any macroscopic index into a trajectory constraint."],"forward_implications":["Synchronization performance becomes a programmable specification: the user sets R* and t*, and the network learns a control to meet them.","The approach can serve as a low-energy control discovery tool in systems where analytic frequency compensation does not exist.","The framework is model-agnostic and extends to amplitude dynamics (Stuart–Landau) and larger networks, as long as a collective observable can be defined.","The observed smoothness of controls may ease actuator constraints, avoiding the saturation issues of phase feedback."],"fun_headline_variants":["Neural net tunes sync timing and strength","Physics-informed network controls sync onset and depth","When and how much to sync: a PINN decides","Trajectory-level control: PINN adjusts sync time and coherence","No feedback law needed: PINN steers oscillator sync"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The hard requirement that R(t) stay at or above R* for all continuous times after t* is only penalized at sampled collocation points, so the network has no guarantee against dips between the samples or in forward simulation.","fun_headline_variants_meta":{"raw":{"variants":["Neural net tunes sync timing and strength","Physics-informed network controls sync onset and depth","When and how much to sync: a PINN decides","Trajectory-level control: PINN adjusts sync time and coherence","No feedback law needed: PINN steers oscillator sync"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1116,"prompt_tokens":641,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":385,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":385,"tokens_out":475,"duration_ms":5028,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T13:07:45.348466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained control law, integrate the deterministic Kuramoto dynamics with a fine ODE solver, and check directly whether R(t) ≥ R* for every t in [t*, T]. If any dip below R* exists, the persistence constraint is violated and the claimed precise regulation fails; a specific probe would test R* = 1, where even tiny phase deviations drop R below 1.","supporting_citations":[],"review_version":1}