{"id":"00769a79-bc7d-4ed8-b36d-4a569ced1af3","arxiv_id":"2505.05521","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A closed-loop neural controller for SPDEs combines regularity-structure features with an operator-encoded policy net, and beats open-loop and RL baselines on two stochastic PDE tracking tasks.","lead":"The paper introduces MB-CC, a neural control algorithm that uses regularity-structured features and a learned policy network to control stochastic partial differential equations. It reports strong tracking performance on reaction-diffusion and Navier-Stokes benchmarks, with closed-loop control computed in a fraction of a second.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported control results may be circular if the policy is evaluated on the same learned model used to train it; the paper never states the evaluation protocol.","rationale":"The reader identified the surrogate-model training loop as the weakest assumption; I agree that it is fragile, but the decisive version is the evaluation protocol. If the policy is trained and evaluated on the same surrogate, the reported improvements are not evidence of real closed-loop control. The paper's Section 4.2 explicitly states that training uses u~_theta rather than the environment, but Section 5 does not describe the evaluation loop. This ambiguity is load-bearing because the headline claim is entirely empirical. I therefore propose a direct check of the code, which is promised in the supplementary materials. Other weaknesses (missing error bars, missing RF hyperparameters n,m,l, and the overstated 'first' claim) are real but secondary. The statistical issue is mitigated by the large margins (e.g., SAC 2.59 vs MB-CC 0.066 on reaction-diffusion), and the missing hyperparameters affect reproducibility but not the truth of the claim. The novelty overstatement concerns framing, not technical soundness. Because the concern is about an unstated protocol rather than a demonstrated contradiction, the appropriate verdict remains CONDITIONAL: the paper should be accepted only after the authors clarify or verify the evaluation protocol. This matches the reader's verdict, so no movement in the verdict.","tokens_in":14460,"tokens_out":9268,"duration_ms":91513,"concrete_test":"Inspect the released code in the supplementary materials to locate the evaluation routine used for Tables 2 and 5. Check whether the tracking objective is computed by rolling out the trained policy on the true numerical SPDE solver (with the same discretization and noise sampling as in data generation) or by using the learned surrogate u~_theta. If the evaluation uses the surrogate, re-run the 1D reaction-diffusion experiment by applying P_gamma to the solver with 50 independent noise samples and compare the resulting objective loss to Table 2; a large discrepancy would confirm the circularity. If the evaluation uses the solver, the circularity concern is resolved, but also compute the surrogate's prediction error on the states visited by P_gamma in closed loop to quantify the model-bias gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MB-CC provides fast, accurate closed-loop control of SPDEs. The policy net P_gamma is trained by backpropagating through the learned surrogate u~_theta (Section 4.2: \"we use the model's prediction u~_theta to simulate the trajectories rather than interact with the environment\"). For this to support the claim, the tracking errors in Tables 2 and 5 must be obtained by applying P_gamma to the true numerical solver of the SPDE, not by re-simulating with u~_theta. The paper never states which protocol is used. If the reported values are computed with u~_theta, then the loss in equation (8) and the evaluation are in-sample for the same model; the improvements over open-loop baselines would only show that the policy exploits the surrogate, and the 'closed-loop' claim would be vacuous because no feedback from the real system is used in testing. Even under the more charitable reading (evaluation on the true solver), the text alone does not establish this, and the missing sim-to-real gap quantification leaves the core quantitative claim unverifiable from the manuscript. This is the load-bearing point: the entire empirical case for MB-CC rests on it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MB-CC, a model-based closed-loop control algorithm for stochastic partial differential equations (SPDEs). The method has two components: a Regularity Feature (RF) block, inspired by regularity structure theory, which transforms the state and forcing into features used by a neural-operator surrogate model; and an operator-encoded policy network Pγ that maps the current state, target state, and time to a control action, trained by backpropagating a tracking/energy objective through the learned surrogate. The method is evaluated on tracking problems for a 1-D stochastic reaction-diffusion equation and a 2-D stochastic Navier-Stokes equation, comparing against open-loop control with CNN/FNO/RF-based surrogates and a Soft Actor-Critic reinforcement learning baseline. The reported results show substantially lower objective loss and much faster inference for the policy-net variants, with ablations across noise scales.","tokens_in":14658,"tokens_out":3356,"duration_ms":35772,"significance":"If the reported results are obtained by evaluating the learned policy on the true numerical SPDE solver, MB-CC would be a meaningful advance in neural closed-loop control for stochastic systems, combining a physically motivated feature block with a fast, feedback-based policy. The paper provides code, evaluates two different backbone architectures to demonstrate plug-and-play behavior, and includes ablations showing robustness to increased noise amplitude. These are concrete strengths. However, the central quantitative claims hinge on an evaluation-protocol detail that the manuscript does not state explicitly, and the surrogate-to-environment gap is treated only as an assumption rather than a measured quantity. The contribution is therefore defensible but requires additional verification before the headline comparisons can be accepted.","major_comments":[{"comment":"The evaluation protocol for the control experiments is not stated. Section 4.2 says that 'throughout the process, we use the model's prediction ũ_θ to simulate the trajectories rather than interact with the environment,' and the policy loss in Eq. (8) is computed through the surrogate. It is therefore unclear whether the objective values reported in Tables 2, 5, and 7 are obtained by rolling out Pγ on the true numerical SPDE solver or by re-simulating with the learned surrogate. If the latter, the improvements over open-loop baselines would be in-sample for the surrogate, and the 'closed-loop control' claim for the actual SPDE would not be supported. The manuscript must state explicitly which environment is used for evaluation and, if the reported numbers are surrogate-based, add an evaluation on the true solver.","section":"§4.2, §5 (Tables 2, 5, 7)"},{"comment":"The load-bearing premise that the learned surrogate is 'close enough to the environment' for policy training and evaluation is not quantified. Forward-model errors in Tables 1 and 4 are reported on the training data distribution, but the policy induces a different state-action distribution, and the tracking objective involves long-horizon rollouts. The paper does not measure how surrogate prediction error propagates to policy loss or closed-loop tracking error. A concrete fix is to compare policy rollouts under the true solver with those under the surrogate, reporting the resulting tracking error and objective value for at least one configuration per equation.","section":"§4.2, §5.1–5.2"},{"comment":"The empirical comparisons lack statistical robustness and baseline-detail information. No error bars, standard deviations, or multiple-seed results are reported for the control metrics. The open-loop baselines are underspecified: the manuscript does not describe how the open-loop control sequence is optimized (e.g., gradient descent through the learned model, number of iterations, initialization, or stopping criterion), which is necessary because the open-loop timing comparisons in Tables 3 and 6 depend on that optimization procedure. The SAC baseline also lacks details such as reward shaping, replay buffer size, environment interface, and hyperparameter tuning. These omissions weaken the strength of the claims that MB-CC outperforms the baselines.","section":"§5.1–5.2, Tables 2–6"}],"minor_comments":[{"comment":"The notation in Eq. (8) is ambiguous: the state is written as ũ_{t,θ}(u0, Pγ(ũ_t, u_T, t), ξ_i), but the argument t appears both as the current time index and as the integration interval endpoint [0,T]; the meaning of ũ_t inside Pγ should be defined more precisely, for instance by writing the time subscript explicitly.","section":"Eq. (8)"},{"comment":"The condition '1 ≤ k + j ≤ m I_{j=0} + ℓ I_{j>0}' in Step 3 is garbled and likely a typesetting artifact; it should be written with explicit indicator functions or a verbal description so that the allowed Taylor orders are unambiguous.","section":"Algorithm 1"},{"comment":"The data augmentation procedure is described only as 'increasing the proportion of difficult-to-fit data in the dataset through duplication.' The paper should specify which samples are considered difficult, the duplication factor, and whether the augmentation is applied to the training set for the forward model, the policy network, or both.","section":"§4.1, Data Augmentation"},{"comment":"The distributions of u0 and f are said to 'follow the previous work [Hwang et al., 2021]' without giving the actual distributions. For reproducibility, the manuscript should state the distributional form or provide the generating code in the supplement.","section":"§5.1"},{"comment":"The horizontal axis is labeled 'Scale of Noise' but the caption and text refer to specific σ values (0.05, 0.2, 0.3, 0.5). The figure should use σ as the axis label or explain the monotone mapping.","section":"Figure 4"},{"comment":"The claim of being 'the first model-based closed-loop control method for SPDEs' is strong; given the prior DDPG-based SPDE control work of Pirmorad et al. [2021] and the existence of closed-loop neural PDE controllers, the novelty statement should be qualified to 'first model-based closed-loop method that combines a learned surrogate with an operator-encoded policy trained by backpropagation' or similar, to avoid overclaiming.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the evaluation of the policy network may have been performed on the learned surrogate rather than on the true SPDE solver, which would make the central empirical claim circular. The authors must clarify this and, if necessary, add a true-environment evaluation. I would also encourage the editor to ask for a small reproducibility package check, since the paper states that code is provided but gives no repository link or verification details beyond the appendix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about arXiv:2505.05521. First, it is a genuine attempt to do closed-loop control of SPDEs with neural operators, combining the regularity-feature (RF) block from the authors' prior line of work with an operator-encoded policy network. Second, the paper never states whether the reported tracking errors are obtained by feeding the policy's actions into the true numerical solver or by re-simulating with the learned surrogate u~_theta. That omission matters more than any other issue.\n\nWhat is actually new: the RF block as a plug-and-play component for different backbones (CNN, FNO), the closed-loop policy that consumes the state and the discretized operator (L_dis u_t), and the demonstration that both pieces help on two benchmark SPDEs. The ablations on noise scale (sigma = 0.05 vs 1) make a decent case that the improvement comes from handling stochasticity, not just from more capacity. The reported speedups (hundreds of seconds to ~0.3s) are large and credible given the policy is a single forward pass.\n\nThe soft spots are real but fixable. Most importantly, the evaluation protocol. Section 4.2 says trajectories are simulated with u~_theta \"rather than interact with the environment\" because it is \"supposed to be close enough.\" That sentence is about training the policy. For the tables in Section 5, the paper is silent on whether the closed-loop control is evaluated on the true SPDE solver. If the errors in Tables 2 and 5 are computed from u~_theta, then the comparison to open-loop baselines is partly in-sample and the \"closed-loop\" claim is vacuous — there is no real feedback from the environment. Even under the charitable reading that they used the true simulator, the text does not say so, and the surrogate-to-real gap is never quantified. This is the load-bearing point, and a referee should insist on an explicit statement plus, ideally, a plot of policy evaluation on both the surrogate and the true solver.\n\nOther issues: no error bars or multiple seeds, so we don't know if the 20–30% improvements are stable; the RF feature heights n,m,l are missing from the hyperparameter tables; the SAC baseline is given no tuning details, making the RL comparison hard to interpret; and the \"first model-based closed-loop control method for SPDEs\" overreaches in the presence of classical stochastic control literature and prior RL work on SPDEs (which they cite). These are minor-to-moderate; none destroys the central idea.\n\nThe circularity worry from our stress-test does not, in my reading, hold if evaluation is on the true simulator — the forward model is trained on simulated data, the policy uses it as a surrogate, and the final metric is against the real dynamics. That is standard. But the ambiguity in the text is enough to keep the paper from being verifiable as written.\n\nWho is this for: anyone working on neural control of PDEs/SPDEs, or on surrogate-based policy learning. It deserves a serious referee. My recommendation: send it to review, but require the evaluation protocol, error bars, and hyperparameters before acceptance.","headline":"The core method is useful and the experiments are extensive, but the paper never states whether the reported control errors come from the true simulator or the learned surrogate, and that needs fixing before I trust the numbers.","tokens_in":15239,"tokens_out":3994,"would_cite":true,"duration_ms":39303,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","93E20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes MB-CC, a closed-loop controller for stochastic PDEs built on regularity-structure features, and reports that it tracks targets faster and more accurately than open-loop neural control or reinforcement learning.","keywords":["stochastic partial differential equations","closed-loop control","regularity structures","neural operator","policy network","reaction-diffusion equation","Navier-Stokes equations","model-based control"],"falsifier":"Run a trained MB-CC policy against a high-accuracy numerical solution of the same SPDE on unseen noise draws and compare the realized tracking objective with the surrogate-evaluated loss; a gap that widens when the noise scale goes from $\\sigma=0.05$ to $\\sigma=1$ would show that surrogate bias, not just stochasticity, limits the closed-loop claim.","tokens_in":14209,"feed_emoji":"🎛️","tokens_out":9296,"duration_ms":83492,"temperature":0.7,"pith_summary":"Stochastic partial differential equations—PDEs driven by random noise—describe turbulence, financial volatility, and other noisy phenomena, but their low-smoothness states make them hard to model and harder still to control. This paper proposes MB-CC, a closed-loop control method made of two pieces: a Regularity Feature block that maps the noisy state and forcing into Taylor-expansion features inspired by regularity structure theory, and an operator-encoded policy network that turns the current state into a control action while treating the learned surrogate as a differentiable simulator. On tracking tasks for the 1-D stochastic reaction-diffusion equation and the 2-D stochastic Navier-Stokes equation, MB-CC reports lower objective loss and much shorter control-computation time than open-loop neural control and the Soft Actor-Critic baseline. If these results hold, closed-loop neural control becomes a practical option for SPDE systems, where optimization-based control is slow and stochasticity destabilizes both learning and control.","feed_headline":"Closed-loop neural controller cuts stochastic-PDE tracking error","feed_subtitle":"A learned surrogate plus feedback policy beats open-loop control and RL on two benchmark SPDE tasks.","key_machinery":"The load-bearing mechanism is the Regularity Feature (RF) block, which generates the feature set $S_{n,m,l}$ by iterating Picard-style integrations of products of the state, its derivatives, and the noise $\\xi$ through the discretized linear operator $(\\mathrm{Id} - L_{\\mathrm{dis}}\\delta t)^{-1}$, mirroring the Taylor expansion of the SPDE's mild solution. These features carry the low-regularity part of the dynamics, and the surrogate output is $\\tilde{u}_\\theta = \\theta_1 s_{\\mathrm{out}} + W_{\\theta_2}(s_{\\mathrm{out}}, O)$, so the backbone network only models the truncation residual. The same discretized operator $L_{\\mathrm{dis}}$ is injected into the operator-encoded policy network by concatenating $u_t$ and $L_{\\mathrm{dis}}u_t$ for the current state and the target state, which gives the feedback law physical information and a single-forward-pass path from state to control.","core_discovery":"The paper presents MB-CC as the first model-based closed-loop control method for SPDEs and claims that its two design choices address the two ways noise breaks control: low regularity breaks the surrogate, and open-loop decisions break the feedback. The RF block computes the Picard/Taylor terms that regularity structure theory singles out—integrals of powers of the state, the deterministic forcing, the noise, and their products against the linear evolution operator—and lets the backbone network fit only the truncated remainder; the operator-encoded policy network observes the current and target states, embeds both with the discretized SPDE operator, and outputs the next external force. The policy is trained by backpropagating the tracking loss through the differentiable surrogate with many sampled noise realizations. The experimental claim is that on the 1-D stochastic reaction-diffusion and 2-D stochastic Navier-Stokes benchmarks this combination lowers the tracking objective and cuts control generation from tens of seconds to a fraction of a second, and that the advantage grows when the noise scale is raised.","pith_inferences":["A natural next test is to add a model-bias or adversarial-disturbance term to the policy loss, since the paper's closed-loop claim is only as strong as the surrogate's fidelity and the paper does not quantify that gap.","The same operator-encoded architecture could be retrained for stabilization, switching setpoints, or energy minimization by changing only the loss functional; the paper demonstrates only tracking, so these remain open extensions.","Because closed-loop control replaces iterative optimization with one network evaluation, the real bottleneck for deployment is the RF-enhanced surrogate's accuracy on out-of-distribution states, which suggests a fine-tuning-on-real-data stage as the decisive practical next step."],"forward_implications":["Control generation becomes a single forward pass: the policy net produces actions in about 0.25–0.31 seconds on the reaction-diffusion task and 0.02–0.10 seconds on the Navier-Stokes task, versus tens of seconds for open-loop optimization.","The RF block is plug-and-play: adding it to a CNN or FNO backbone cuts forward-modeling error on the reaction-diffusion equation from 0.0182 to 0.0115 for CNN and from 0.0138 to 0.0029 for FNO.","Every backbone tested works better with the policy net than in open-loop mode, and the full MB-CC combination beats SAC by at least an order of magnitude in objective loss on both benchmark SPDEs.","Raising the noise scale from $\\sigma = 0.05$ to $\\sigma = 1$ degrades open-loop control sharply, while the RF-policy combination holds its objective approximately constant, supporting the claim that the gains come from stochasticity handling."],"supporting_citations":[{"why":"Supplies the regularity structure theory whose Picard/Taylor terms the RF block turns into features.","marker":"[Hairer, 2014]"},{"why":"Earlier regularity-structure neural operator for SPDEs that MB-CC's RF block parameterizes and extends.","marker":"[Hu et al., 2022]"},{"why":"Shows regularity features improve SPDE modeling, providing the line of work that the RF block builds on.","marker":"[Gong et al., 2023]"},{"why":"Provides the feature-engineering justification and the numerical setup used to generate noise and data for the reaction-diffusion equation.","marker":"[Chevyrev et al., 2024]"},{"why":"Establishes the open-loop surrogate-model control baseline that the closed-loop policy net is compared against and improves on.","marker":"[Hwang et al., 2021]"},{"why":"Neural SPDE model and numerical solver used to generate the stochastic Navier-Stokes data.","marker":"[Salvi and Lemercier, 2021]"},{"why":"Fourier Neural Operator serves as one backbone for testing the RF block's plug-and-play behavior.","marker":"[Li et al., 2021]"},{"why":"Soft Actor-Critic is the reinforcement-learning baseline that MB-CC outperforms in the tracking experiments.","marker":"[Haarnoja et al., 2018]"}],"fun_headline_variants":["First model-based closed-loop control for stochastic PDEs","MB-CC: tackling SPDE control with regularity features and feedback","Neural operator + regularity theory = robust SPDE control","Stochastic PDE control: new algorithm beats open-loop and RL","Closed-loop SPDE control made fast and robust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The policy network is trained entirely on trajectories predicted by the learned surrogate model rather than on the true SPDE, so the controller's quality depends on the surrogate being close enough to the real system in exactly the state regions the policy visits.","fun_headline_variants_meta":{"raw":{"variants":["First model-based closed-loop control for stochastic PDEs","MB-CC: tackling SPDE control with regularity features and feedback","Neural operator + regularity theory = robust SPDE control","Stochastic PDE control: new algorithm beats open-loop and RL","Closed-loop SPDE control made fast and robust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000774,"raw_usage":{"total_tokens":3483,"prompt_tokens":1060,"completion_tokens":2423,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":2342}},"tokens_in":676,"tokens_out":2423,"duration_ms":19972,"temperature":1.0,"reasoning_tokens":2342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:17:52.163231+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a trained MB-CC policy against a high-accuracy numerical solution of the same SPDE on unseen noise draws and compare the realized tracking objective with the surrogate-evaluated loss; a gap that widens when the noise scale goes from $\\sigma=0.05$ to $\\sigma=1$ would show that surrogate bias, not just stochasticity, limits the closed-loop claim.","supporting_citations":[{"cited_title":"Deep latent regularity network for modeling stochastic partial differential equations","cited_arxiv_id":null,"evidence_quote":"Shows regularity features improve SPDE modeling, providing the line of work that the RF block builds on."},{"cited_title":"Feature engineering with regularity structures","cited_arxiv_id":null,"evidence_quote":"Provides the feature-engineering justification and the numerical setup used to generate noise and data for the reaction-diffusion equation."},{"cited_title":"Solving pde-constrained control problems using operator learning, 2021","cited_arxiv_id":null,"evidence_quote":"Establishes the open-loop surrogate-model control baseline that the closed-loop policy net is compared against and improves on."}],"review_version":1}