{"id":"2426c04d-0927-4b66-985e-8ed840a748c6","arxiv_id":"2509.23975","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural-operator surrogate plus Newton-Krylov and Arnoldi methods yields a reduced model for LQR and pole-placement controllers, but the closed-loop demonstration is performed on the surrogate under an assumed additive actuation model, not on the true PDE.","lead":"The authors combine local neural operator surrogates with matrix-free Krylov methods to design stabilizing controllers for a model PDE, replacing the microscopic simulator with a learned timestepper. The paper's own validation, however, tests the controller on the surrogate model itself, not on the true physics, and the abstract overstates the mismatch analysis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop validation only exercises the additive surrogate (Eq. 25) used to design the gain; the NO-designed controller is never applied to the true controlled PDE, so the abstract's claim of consistent true-system performance is unsupported.","rationale":"The reader's weakest_assumption is exactly the load-bearing concern: the controlled timestepper is approximated by u_{n+1} ≈ S_Δt(u_n) + B z_n (Eq. 25), with B taken directly from the PDE and no higher-order Duhamel terms, and the closed-loop test uses the same approximation. The paper's own limitation statements confirm this reading: 'we do not have access to a fully trained controlled timestepper', 'the RandONet-based controllers were designed and tested under the assumption that the control action is applied in the approximated actuator effect form, and not to the actual physical model', and 'the framework's robustness to plant-model mismatch remains to be investigated in future work.' These are explicit, in-scope admissions that the central validation is missing. The abstract overclaims by saying 'consistent performance between the learned surrogate and the true system' is demonstrated and that degradation under plant-model mismatch is quantified. Neither is supported: consistency is between two differently designed controllers applied to their own models, and mismatch robustness is deferred. This is a serious but addressable gap, not a fatal flaw. The proposed pipeline is coherent and follows established EF/Krylov techniques; the RandONet surrogate is reused from prior work; and the controller design itself is standard once (F,D) are available. The missing experiment is straightforward: apply the NO-designed gain to the high-fidelity controlled PDE and check closed-loop stability. If that experiment succeeds, the central claim is substantially vindicated; if it fails, the additive actuation approximation or the surrogate's Jacobian/Arnoldi identification is inadequate. Thus the reader's conditional verdict is appropriate and unchanged.","tokens_in":12859,"tokens_out":4361,"duration_ms":70367,"concrete_test":"Re-run the closed-loop experiment of Fig. 2(d) with the RandONet-designed K (dLQR and PP) applied to the FD/Euler controlled PDE simulator Φ_FD(u_n,z_n) with inner step Δt_FD=0.0001, same u_0 and same feedback z_n = -K V_M^T(u_n - u_ss) at each outer step Δt=0.001. The concern is settled if ||u(t)-u_ss||_2 decreases below a physical tolerance (e.g., <1e-3) and remains bounded; if it diverges or oscillates, the additive actuation model of Eq. 25 is not adequate and the central claim fails. If no FD simulator is available, use a differently discretized PDE solver (e.g., Crank-Nicolson with Δt=0.0005) to break the surrogate self-consistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a controller designed purely from the learned autonomous timestepper stabilizes the real distributed-parameter system. The validation does not show this. In Section IV, the controlled surrogate is taken as u_{n+1} ≈ S_Δt(u_n) + B z_n (Eq. 25), explicitly neglecting Duhamel/higher-order terms. The reduced model (F,D) and the dLQR/PP gains are based on this first-order additive actuation. Figure 2(d) reports convergence for 'both FD model and RandONet', but the RandONet curves are generated by applying the designed control to this same approximate surrogate, not to the true controlled PDE. The FD benchmark is a separate controller designed from the PDE-based timestepper. The paper itself states that RandONet-based controllers 'were designed and tested under the assumption that the control action is applied in the approximated actuator effect form, and not to the actual physical model' and that robustness to plant-model mismatch 'remains to be investigated in future work.' Because the surrogate includes the exact B used in the controller, the closed-loop test is partially circular: it exercises the same model used to synthesize the gain. It therefore cannot substantiate the abstract's claim of 'consistent performance between the learned surrogate and the true system, with quantified degradation under plant-model mismatch.' The pipeline may well work, but the missing experiment—closed-loop simulation of the true controlled PDE using the NO-designed gains—is the load-bearing evidence needed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven, equation-free control pipeline for distributed parameter systems. A local RandONet surrogate, trained on short-horizon autonomous simulation data, replaces the fine-scale timestepper. Newton-Krylov GMRES is used to compute unstable steady states, and Arnoldi iterations provide a low-dimensional slow subspace and a reduced linearized model (F,D). Discrete-time LQR and pole-placement controllers are designed on this reduced model and lifted to the full nonlinear timestepper. The method is demonstrated on the Liouville-Bratu PDE with three Gaussian actuators, comparing RandONet-based controllers with finite-difference-based controllers. The core limitation, acknowledged in Section IV-A, is that the RandONet controllers are tested only on the approximate additive-actuation surrogate u_{n+1}=S_Δt(u_n)+Bz_n, not on the actual controlled PDE; the abstract's claim of 'consistent performance between the learned surrogate and the true system' is therefore not supported by the reported experiments.","tokens_in":13241,"tokens_out":4026,"duration_ms":33010,"significance":"If fully validated, the approach would be a useful contribution: it extends local neural-operator surrogates from trajectory generation and bifurcation analysis to coarse feedback control, and it uses only short-horizon autonomous data plus a prescribed actuation model. The matrix-free Newton-Krylov and Arnoldi components are standard and appear correctly applied. The paper is honest in stating that actuator effects are not learned from data and that robustness to plant-model mismatch is future work. However, the end-to-end claim is currently stronger than the evidence: the closed-loop test for the RandONet controller exercises the same additive surrogate used to synthesize the gain, so the stabilization result is partly circular. The significance of the paper would increase substantially if the NO-designed gain were applied to the true controlled PDE (e.g., the FD timestepper with inner integration) or if actuator effects were learned from controlled data.","major_comments":[{"comment":"The central validation is missing. The RandONet closed-loop trajectories are generated using u_{n+1}=S_Δt(u_n)+Bz_n, with B prescribed from the PDE. The only controller applied to the actual controlled PDE is the FD-based one. Consequently, Fig. 2(d)'s comparison between 'RandONet' and 'FD' does not establish that the NO-designed controller stabilizes the true physical system. The manuscript itself states that the RandONet controllers 'were designed and tested under the assumption that the control action is applied in the approximated actuator effect form, and not to the actual physical model.' This directly contradicts the abstract's claim of 'consistent performance between the learned surrogate and the true system.' Please add a closed-loop simulation in which the RandONet-designed gain is applied to the exact controlled timestepper (e.g., the FD model with inner Δt_FD=0.0001), or remo","section":"Section IV-A, Eq. (25), Fig. 2(d)"},{"comment":"The claim of an 'end-to-end equation-free control pipeline where the only requirement is a dataset of short-horizon simulations' is overstated. The actuator structure B in Eq. (25) is taken from the known PDE, and the paper explicitly says 'we do not learn the exact actuator effect from data.' Thus the pipeline still requires knowledge of the actuation mechanism. To support the data-only claim, the actuator effect should be learned or inferred from data, or the claim should be restricted to replacement of the autonomous timestepper. This is a load-bearing issue for the stated novelty.","section":"Introduction, Section IV-A, Conclusion"},{"comment":"The abstract promises 'quantified degradation under plant-model mismatch,' but no plant-model mismatch experiment is reported. The only related observation is that the RandONet controller saturates at an error of about 10^-7 instead of machine precision, which is a surrogate-accuracy limitation, not a quantification of closed-loop robustness to mismatch. The paper also states that robustness 'remains to be investigated in future work.' Please either add mismatch experiments (e.g., perturbing B or the surrogate) or remove this claim from the abstract.","section":"Abstract and Section IV-A"}],"minor_comments":[{"comment":"The definition of D is under-specified. V_F is said to contain eigenvectors of F, but Eq. (17) uses the same F for y_{n+1}=F y_n. Please clarify whether y_n are coordinates in the Arnoldi basis or in the eigenbasis of F, and define H explicitly (derivatives of S_T with respect to the actuators at the steady state). Currently the dimensions and the role of V_F are ambiguous.","section":"Eq. (19)"},{"comment":"The symbol F is overloaded: it denotes the general operator being approximated in Eq. (5) and later the reduced Jacobian matrix in Eq. (17). This is a readability issue; consider renaming one of them.","section":"Section II-B and Section III-B"},{"comment":"The phrase 'coarse stable and unstable steady states' and 'steady and unsteady-states' should be unified and corrected; 'unsteady' usually means time-dependent, not unstable.","section":"Abstract and Introduction"},{"comment":"Panel (e) is described as 'dLQR RandONet convergence' in the caption text but the caption itself says 'Spatiotemporal evolution.' Please make the caption and panel labels consistent.","section":"Fig. 2 caption"},{"comment":"Minor typo: 'spatially structureed' should be 'spatially structured.' Also, the sentence beginning 'An important modeling challenge arises...' appears in the Conclusion but reads like a Discussion section; consider moving it before the closing paragraph.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound methodological core, but its abstract and conclusion are materially stronger than the evidence. The missing experiment—applying the NO-designed gain to the true controlled PDE—is the key gate for acceptance. The authors are transparent about this gap in Section IV-A, so this is an overclaim rather than an omission that misleads by silence. If the missing closed-loop test can be supplied and confirms stabilization, or if the claims are appropriately downgraded, the paper would be publishable. I would not recommend rejection, since the framework and derivations are credible and the missing validation is within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a legitimate, workmanlike extension of the equation-free control toolbox: instead of calling a microscopic simulator on demand, the authors train a RandONet surrogate on short-horizon data and then run the standard Newton-Krylov/Arnoldi machinery through it to get a reduced model for dLQR or pole-placement design. That specific combination is new, even though the ingredients come from prior work. The paper is clearly written and the numerical pipeline is standard, so the math is fine as far as it goes. The FD benchmark is a sensible sanity check, and the authors are upfront about the additive actuation approximation and the fact that robustness is future work. Those honest admissions are a point in the paper's favor.\n\nThe soft spot is the one the stress-test note identifies, and it is load-bearing. The abstract claims consistent performance between the learned surrogate and the true system, but the RandONet controller is never applied to the true controlled PDE. Equation (25) writes the controlled surrogate as S_dt(u_n)+B z_n, and the closed-loop simulations in Figure 2 use that same model. So the controller is designed and tested on the same first-order additive approximation. The FD controller, by contrast, is applied to the actual PDE with properly integrated actuator effects. Comparing those two is not a fair or meaningful demonstration of plant-model mismatch. The conclusion's \"first time\" claim is also a bit strong given earlier EF control work with actual simulators.\n\nThese are fixable, not fatal. The cleanest fix is to take the RandONet-designed gains and apply them to the true PDE in closed loop (or a more accurate controlled timestepper) and report the resulting error. If that experiment works, the surrogate-only testing becomes a minor caveat rather than a central hole. If the authors prefer not to run it, they should scale back the abstract and conclusion to say the controller stabilizes the surrogate model under an assumed actuator form, with true-system validation left as open work. Missing code and data also hurt reproducibility, though that is easier to remedy.\n\nWho gets value from this: researchers in equation-free methods, operator-learning for control, and anyone who wants a reference for how to wrap a neural operator in matrix-free linear algebra. It deserves a serious peer review, but the review should push for the missing true-system experiment or a revision that matches claims to evidence.","headline":"Useful extension of equation-free control to neural-operator surrogates, but the central claim that the controller works on the true system is unsupported: it is only tested on the same approximate surrogate used to design it.","tokens_in":13739,"tokens_out":1391,"would_cite":true,"duration_ms":13984,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C20","68T07","65H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a local neural operator trained only on short-horizon data can replace a first-principles timestepper in equation-free control, demonstrated by stabilizing an unstable Liouville–Bratu steady state.","keywords":["equation-free methods","neural operators","local-in-time surrogates","coarse control","distributed parameter systems","Krylov subspace methods","discrete-time LQR","Liouville–Bratu equation"],"falsifier":"Take the gain matrix K produced by the neural-timestepper reduced model and apply it to a controlled PDE integrator that evaluates actuator influence through the exact Duhamel integral (or a sufficiently fine inner-step integration), rather than the additive B z_n update. If the closed-loop L2 error to the target steady state does not decay, the end-to-end claim is refuted.","tokens_in":12747,"feed_emoji":"🎛️","tokens_out":7016,"duration_ms":47025,"temperature":0.7,"pith_summary":"The paper tries to show that you can design a stabilizing feedback controller for a high-dimensional, spatially distributed system without ever writing down its coarse-grained equations or running a microscopic simulator on demand. Instead, a neural operator trained on short-horizon spatiotemporal data becomes the timestepper, and matrix-free Krylov methods extract the steady state, the slow subspace, and a reduced linear model from it. Standard discrete-time control design (dLQR or pole placement) then produces a gain that is applied to the full nonlinear surrogate. The demonstration on the Liouville–Bratu PDE shows the surrogate-based controller driving the system to an unstable steady state with behavior consistent with a controller designed from the known PDE. If this holds generally, it would let controllers be built from data alone for systems where first-principles simulators are unavailable or too expensive.","feed_headline":"Stabilizes unstable PDE state with data-only neural timestepper","feed_subtitle":"Local neural operator replaces the PDE solver in equation-free feedback control of an unstable steady state.","key_machinery":"The load-bearing object is the learned short-time solution operator S_Δt (a local neural operator, implemented with a random-projection architecture), used as a matrix-free oracle for directional derivatives. Around that operator, Newton–Krylov GMRES locates the unstable steady state and Arnoldi iteration exposes the dominant slow modes; the reduced open-loop model y_{n+1}=F y_n + D z_n then lets standard discrete-time control tools (dLQR, pole placement) produce a gain that is lifted to the full nonlinear timestepper.","core_discovery":"The paper's central claim is that a local neural operator, trained on short-horizon spatiotemporal snapshots of an autonomous distributed parameter system, can replace the microscopic/first-principles timestepper in the equation-free methodology. Around this surrogate, the paper runs Newton–Krylov GMRES to locate an unstable coarse steady state, Arnoldi iteration to obtain its dominant Ritz spectrum and slow subspace, and actuator derivative estimates to build a reduced discrete-time model y_{n+1}=F y_n + D z_n. On this reduced model, discrete-time LQR and pole-placement gains are designed and then lifted back to the full nonlinear neural timestepper. For the Liouville–Bratu PDE at λ=2, both","pith_inferences":["The paper leaves untested how the gain behaves on a true controlled integrator where actuation enters through the semigroup (Duhamel form) rather than the additive B z_n term; a natural next step is to apply the designed gain to such a simulator and measure the remaining stabilization.","Because the reduced model's dimension equals the number of retained Ritz modes, the approach is most plausible when the actuators' influence is well captured by the dominant slow subspace; systems with strong coupling between actuators and fast, stable modes may require extending the subspace or an output-feedback variant.","The pipeline suggests a general template: any matrix-free coarse-level task that traditionally wraps a fine-scale simulator—bifurcation continuation, sensitivity analysis, observer design—can in principle be rerun around a neural timestepper learned from data, extending the paper's control result to a broader equation-free toolbox."],"forward_implications":["Controllers for high-dimensional distributed parameter systems can be designed without explicit equations, using only short-horizon trajectory data.","The same learned timestepper serves all three stages of the pipeline—fixed-point finding, stability analysis, and control design—so no on-demand simulation calls are needed.","Both dLQR and pole-placement gains computed from the reduced slow model stabilize the full nonlinear surrogate in the Liouville–Bratu example.","Closed-loop performance is limited by surrogate accuracy: the neural-timestepper controller reaches an L2 error around 1e-7, versus near machine precision for the exact-model controller."],"fun_headline_variants":["Data-only neural timestepper stabilizes unstable PDE","Neural operator replaces solver to control unstable PDE","Surrogate timestepper enables equation-free PDE control","Data-driven coarse control stabilizes unstable PDE steady state","Neural surrogate tames unstable PDE via data-only timestepper"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the controlled timestepper is well approximated by u_{n+1} ≈ S_Δt(u_n) + B z_n; if the true actuator dynamics differ, the designed gain may fail on the physical system.","fun_headline_variants_meta":{"raw":{"variants":["Data-only neural timestepper stabilizes unstable PDE","Neural operator replaces solver to control unstable PDE","Surrogate timestepper enables equation-free PDE control","Data-driven coarse control stabilizes unstable PDE steady state","Neural surrogate tames unstable PDE via data-only timestepper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000825,"raw_usage":{"total_tokens":3456,"prompt_tokens":765,"completion_tokens":2691,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":2613}},"tokens_in":509,"tokens_out":2691,"duration_ms":19341,"temperature":1.0,"reasoning_tokens":2613,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:37:24.815159+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the gain matrix K produced by the neural-timestepper reduced model and apply it to a controlled PDE integrator that evaluates actuator influence through the exact Duhamel integral (or a sufficiently fine inner-step integration), rather than the additive B z_n update. If the closed-loop L2 error to the target steady state does not decay, the end-to-end claim is refuted.","supporting_citations":[],"review_version":1}