{"id":"d3d19857-302e-44c9-8ab5-cc44cbcfbd0f","arxiv_id":"2412.09942","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A learned reduced-order feedback controller computes near-optimal distributed controls for parametrized PDEs in real time, with a latent loop that works even without online state measurements.","lead":"This paper trains neural networks to instantly compute optimal control actions for high-dimensional physical systems, replacing slow full-order simulations. It tests the method on two optimal transport problems, showing speedups of more than a thousand times while keeping the state close to its target.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The policy is trained open-loop and evaluated closed-loop; no evidence that policy-induced states stay on the training manifold, so the real-time control claim is not established.","rationale":"The reader's weakest assumption correctly identifies the distribution shift between open-loop training pairs and closed-loop application, and my reading strengthens it with a specific technical detail: the latent forward model φN is trained and evaluated only one step ahead, while the latent feedback loop uses it recursively without any error-bounding mechanism. This is a load-bearing gap because the abstract promises real-time control for any observed state and the latent loop promises continuous control without full-order measurements, but the reported 4–7% errors are per-snapshot and not closed-loop. The numerical demonstrations in Figures 9 and 14 are suggestive and I credit the authors for testing new initial configurations and scenarios, yet those are single visual examples rather than quantitative closed-loop statistics. My concern is not that the method is wrong, but that the evidence as presented does not yet establish the strongest claims. The reader's CONDITIONAL verdict is appropriate; no change is needed. A concrete closed-loop rollout benchmark with variance estimates would settle the matter.","tokens_in":26248,"tokens_out":6068,"duration_ms":84544,"concrete_test":"Run a closed-loop benchmark on held-out test scenarios: for each test initial-state/scenario pair, simulate the FOM with the learned policy applied at every time step and record final L2 distance to target and closed-loop cost; then repeat with the latent feedback loop unrolled from only the initial state, measuring how one-step φN errors compound over Nt steps. Compare against the open-loop optimal cost and target error. If mean closed-loop error remains near the 7% policy error and does not grow with horizon, the concern is mitigated; if it degrades materially or is reported without variance across scenarios, the real-time generalization claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'optimal control actions are retrieved in real-time for any observed state and scenario' and that the latent loop can 'continuously control the system'—rests on the unverified assumption that πN, trained on open-loop optimal state-control pairs via Eq. (5), remains accurate when applied iteratively to states it itself generates. Section 3.2 simply loops forward passes, and Section 3.3 replaces missing states by recursive evaluation of φN, yet JφN in Eq. (7) is only a sum of one-step prediction errors on training states. There is no bound, no compounding-error experiment, and no distribution-shift analysis for policy-induced rollouts. The reported 4–7% per-snapshot errors therefore do not by themselves imply closed-loop target-reaching. The Conclusions' assertion that the method avoids 'losses of optimality, performance or stability' is unsupported by the presented evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an offline-online reduced-order feedback control strategy for PDE-constrained optimal control problems with distributed state and control variables. During the offline phase, optimal state-control trajectories are generated for sampled initial conditions and scenario parameters using the adjoint method; the snapshots are reduced via POD followed by an autoencoder (POD+AE); and a feedforward policy is trained to map low-dimensional state and scenario parameters to low-dimensional controls. An optional latent forward model is trained and used to close the loop when full-order state measurements are unavailable. Two Fokker-Planck optimal transport problems (in vacuum and in a fluid with an obstacle) are used to evaluate reconstruction, one-step policy prediction, forward-model prediction, noise robustness, and online speedups, with reported per-snapshot errors around 4–7% and speedups of 1000×–32000×.","tokens_in":26389,"tokens_out":4706,"duration_ms":53981,"significance":"The main potential contribution is a non-intrusive, real-time feedback control pipeline for high-dimensional parametrized PDEs, combining POD+AE reduction with a learned policy and a latent forward model. The numerical experiments are nontrivial and include a noise robustness study over 100 scenarios. If the closed-loop performance were demonstrated against full-order optimal feedback benchmarks, this would be a practical and useful advance over MPC and DRL baselines. However, the current evidence is primarily one-step supervised errors on test snapshots and a few qualitative closed-loop trajectories; the claims of 'optimal' control and of preservation of stability and performance exceed what is demonstrated.","major_comments":[{"comment":"The trained policy πN and forward model φN are evaluated only through one-step supervised losses on test snapshots, while the online phase in Section 3.2 and the latent loop in Section 3.3 apply them recursively to states produced by the policy itself. No analysis or experiment addresses the resulting distribution shift or the compounding of forward-model errors over Nt steps; the reported 4–7% errors are therefore not evidence for the abstract's claim that 'optimal control actions are retrieved in real-time for any observed state', nor for the conclusion's claim that the latent loop avoids 'losses of optimality, performance or stability'. Please provide closed-loop rollouts for multiple unseen scenarios, quantify the distance from the full-order optimal closed-loop trajectory or cost, and report multi-step latent prediction error (e.g., error after k steps without resetting).","section":"Section 3.2 and 3.3, Eqs. (5), (7)"},{"comment":"The speed-up comparison is not against a full-order feedback controller. The online times are compared with the time to solve a single open-loop OCP (15 or 22 minutes per scenario), whereas a true full-order feedback loop would require repeated OCP solves at every time step. Please clarify the baseline and, ideally, compare the closed-loop performance and cost with a receding-horizon FOM controller for at least one scenario.","section":"Section 4.1 and 4.2 (speed-up claims)"},{"comment":"All reported errors (e.g., εu_rel = 7.08% in Test 1.2) come from a single train-test split and a single initialization. Since the data generation and network initialization are random, the paper should report variance over several independent runs or data splits; otherwise the claimed accuracy levels and the noise-robustness conclusions cannot be assessed for statistical significance.","section":"Section 4.1 and 4.2 (statistical evaluation)"}],"minor_comments":[{"comment":"The phrase 'Fokker-Plank' should be corrected to 'Fokker-Planck'.","section":"Section 4.2"},{"comment":"The definitions of εy_rel and εu_rel do not specify whether the norm is the Euclidean norm and which time instants are included; please clarify the notation.","section":"Section 4 (mean relative errors)"},{"comment":"The after-decoding error term uses φu_D on both the target and the prediction; consider writing the prediction more explicitly as φu_D(πN(y_N, µ_s)) to avoid ambiguity.","section":"Eq. (5)"},{"comment":"The closed-loop results are presented only qualitatively; please add quantitative metrics (e.g., ||y(T)-y_d|| or arrival probability) for these specific test cases.","section":"Figures 9 and 14"},{"comment":"The index set I'_train is introduced but not fully characterized; please state whether it includes all intermediate time steps or only a subset.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"This is a reasonable engineering contribution to the ROM-for-control literature, but the abstract and conclusions overstate what the evidence supports. The requested closed-loop evaluation experiments are necessary to justify the 'optimal' and 'stability' claims. I would also encourage the editors to ask the authors to make code or data available, as the current manuscript does not appear to ship reproducible artifacts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth taking seriously. It extends the authors' previous open-loop DL-ROM controller to closed-loop feedback and adds a latent forward model that closes the loop when state measurements are missing. That combination—offline optimal-control data, POD+AE compression, learned policy, latent state surrogate—is not in the prior literature, and the two optimal-transport test problems are genuinely high-dimensional and transport-dominated. Full-order noisy-data robustness stress test is a nice addition, and the paper is honest about the offline cost.\n\nWhat it does well: the pipeline is cleanly described, the split into offline generation/reduction/training and online forward passes is coherent, and the reported speedups (1000x/32000x on test 1, etc.) are credible as measurement. The paper makes explicit that only synthetic full-order data are used, and the noise experiment is handled only online, so those claims are appropriately scoped.\n\nSoft spots, in order of size. First, the central generalization claim—'optimal control actions are retrieved in real-time for any observed state and scenario'—is supported by one-step prediction errors and by visual target-reaching in a few curated rollouts. There is no quantitative comparison of policy-induced closed-loop trajectories against the true optimal ones, and no variance estimate from repeated train/test splits. The stress-test worry about distribution shift is fair: the policy is trained on optimal open-loop pairs, and while the paper does simulate online rollouts, it never measures whether policy-induced states stay where the policy is accurate. That is the main load-bearing gap. Second, the latent forward model is evaluated only one-step-ahead; the recursive latent loop in Section 3.3 is applied without a compounding-error analysis. The claim in the Conclusions that the method avoids 'losses of optimality, performance or stability' is stronger than anything demonstrated—the authors don't measure closed-loop optimality or stability at all. Third, no code/data are released, and the single 80/20 or 90/10 split means the reported 4-7% errors could move nontrivially.\n\nNone of this invalidates the method; it means the paper currently demonstrates a promising pipeline, not a certified controller. For a methods paper at the ROM/control boundary, that is a normal and fixable state. The fix is straightforward: run multiple splits, compare policy-driven trajectories to the optimal ones on held-out scenarios, and report compounding error for the latent loop over Nt steps.\n\nFor a reader working on data-driven feedback control or reduced-order modeling, this is useful and should be sent to serious peer review. It needs revision, not rejection.","headline":"A useful extension of the authors' open-loop DL-ROM controller to feedback and latent-loop settings, with solid numerical evidence but no closure guarantee or artifact release.","tokens_in":26951,"tokens_out":2535,"would_cite":true,"duration_ms":31695,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K20","93B52","68T07","65M60"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a deep-learning reduced-order feedback controller, trained offline on adjoint-generated optimal trajectories, retrieves near-optimal controls in real time from any observed state and scenario, and can keep…","keywords":["feedback control","PDE-constrained optimal control","deep learning-based reduced order models","parametric systems","proper orthogonal decomposition","autoencoders","latent feedback loop","optimal transport"],"falsifier":"One concrete check: deploy the trained full-order feedback controller on a test scenario not used in training, apply each predicted control to the high-fidelity PDE, and compute the probability of arrival $P(Y(T)\\in B_{0.5}(\\mu^d_1,\\mu^d_2))$; if this probability falls well below the noise-free test levels for both the full-order loop and the latent loop, or if the latent-loop trajectory diverges from the full-order-feedback trajectory while the one-step forward errors remain at the paper's reported few-percent levels, the central claim fails.","tokens_in":26009,"feed_emoji":"🎛️","tokens_out":9741,"duration_ms":103321,"temperature":0.7,"pith_summary":"The paper tries to establish that optimal feedback control of high-dimensional distributed systems governed by parametrized PDEs can be made real-time by moving the expensive computation offline. The proposed pipeline generates optimal state-control trajectories with an adjoint-based solver, compresses states and distributed controls into small latent coordinates using POD combined with a deep autoencoder, and trains a feedforward network to approximate the map from current state and scenario to optimal control. Once trained, the control for a new state and scenario is a few forward passes through small networks, with no online optimization. The same latent space also contains a learned forward model, so if full-order measurements are missing the controller can keep issuing controls from predicted latent states. On two two-dimensional optimal-transport problems the resulting feedback laws reach the target with control errors around seven percent and run roughly one thousand to tens of thousands of times faster than full-order solves.","feed_headline":"Latent loop drives transport to target 1,000x faster","feed_subtitle":"After offline training, a learned policy steers high-dimensional densities with ~7% error and over 1,000x speedups.","key_machinery":"The machinery is a two-stage reduction combined with joint training. State and control snapshots are first projected onto POD modes and then encoded and decoded by deep autoencoders, yielding latent coordinates $y_N=\\varphi^y_E(V_y^\\top y_h)$ and $u_N=\\varphi^u_E(V_u^\\top u_h)$; the surrogate policy $\\pi_N$ is a feedforward network mapping latent state and scenario parameters to latent control. When measurements are missing, a second feedforward network $\\varphi_N$ maps $(y_N,u_N,\\mu_s)$ to the next latent state. All networks are trained together by minimizing one cumulative loss that combines state and control reconstruction errors, the policy's latent-space and after-decoding prediction errors, and, in the latent-loop variant, one-step forward prediction errors computed from both data and policy outputs. This joint objective is what makes the latent coordinates meaningful for control rather than for reconstruction alone, and it is the mechanism that lets a policy in ten to eighteen dimensions replace a high-dimensional optimal-control solve.","core_discovery":"The central claim is that the full-order policy $\\pi_h: \\mathbb{R}^{N^y_h}\\times\\mathcal{P}\\to\\mathbb{R}^{N^u_h}$ can be replaced, for online purposes, by a latent policy $\\pi_N: \\mathbb{R}^{N_y}\\times\\mathcal{P}\\to\\mathbb{R}^{N_u}$ trained on optimal state-control pairs produced by the adjoint method. The paper shows that after generating $N_s$ optimal trajectories offline, a POD-plus-autoencoder reduction to latent dimensions around $N_y\\approx 10$–$14$ and $N_u\\approx 14$–$18$, followed by a small feedforward network, predicts the distributed velocity-field control for unseen initial states and scenarios with full-order relative errors around $7\\%$. It further claims that a concurrently trained forward map $\\varphi_N$ at the latent level can close the feedback loop when full-order measurements are unavailable, so control continues from predicted latent states rather than measured ones. Together these pieces make real-time feedback control of high-dimensional parametrized transport-dominated systems a matter of a few forward passes through light neural networks.","pith_inferences":["A natural extension is to monitor the latent forward model's error online and switch back to full-order feedback when its prediction drifts; the paper does not provide such an error-bounding mechanism.","Because the policy is trained on optimal trajectories sampled from chosen initial-state and scenario distributions, active or adaptive sampling in the offline phase would be the direct test of whether the method resists distribution shift when the closed-loop states differ from the training states.","The encoder could be retrained on partial observations $z_h = O_h y_h$, turning the same architecture into a sensor-based controller for camera or sparse-sensor data, which the paper lists as a future direction.","The training-data generation remains expensive, so combining a few high-fidelity trajectories with many cheap surrogate-generated ones—an idea the paper mentions for future work—would determine whether the offline cost can be amortized to practical levels."],"forward_implications":["If the central claim is correct, optimal closed-loop control for these systems no longer requires solving an optimization problem online; each control step is a forward pass through small encoder, policy, and decoder networks.","The latent feedback loop would let a controller keep issuing controls after sensor loss or during measurement latency, since the forward model supplies the next state from the control it just issued.","A single trained architecture covers multiple scenarios encoded in $\\mu_s$—different target locations and, in the fluid test, different inflow speeds and angles—so new cases are handled without retraining.","Because the state and control autoencoders discard high-frequency components, the policy should remain usable with noisy state measurements, as demonstrated by the near-target arrival probabilities in the noise experiments."],"supporting_citations":[{"why":"Supplies the constrained OCP formulation, cost functionals, and KKT/adjoint optimality system used to generate training snapshots.","marker":"[50]"},{"why":"Establishes the earlier open-loop real-time control pipeline that this paper extends with state feedback and latent loop closure.","marker":"[72]"},{"why":"Provides the POD-DL-ROM (POD plus autoencoder) reduction strategy that the paper adopts for state and control compression.","marker":"[27]"},{"why":"Introduces the nonlinear DL-ROM autoencoder architecture whose encoder-decoder structure and joint training the approach builds on.","marker":"[26]"},{"why":"Supplies the POD-NN non-intrusive ROM idea of replacing full-order maps by neural networks trained on reduced snapshots.","marker":"[33]"},{"why":"Defines model predictive control, the surrogate-model closure concept that motivates the paper's latent forward model.","marker":"[22]"},{"why":"Adjoint-based optimal-control solver used to compute the full-order optimal trajectories in the data-generation phase.","marker":"[52]"}],"fun_headline_variants":["Latent feedback loop drives transport 1,000x faster","Learned latent policy controls transport with 7% error","Deep reduced-order model enables real-time distributed control","POD-autoencoder policy speeds closed-loop control by 1000x","Latent policy predicts velocity-field control for new scenarios"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the online closed loop stays inside the training distribution: states reached under the policy's own controls must resemble the optimal trajectories on which $\\pi_N$ and $\\varphi_N$ were trained, and the forward model's one-step errors must not compound over the control horizon.","fun_headline_variants_meta":{"raw":{"variants":["Latent feedback loop drives transport 1,000x faster","Learned latent policy controls transport with 7% error","Deep reduced-order model enables real-time distributed control","POD-autoencoder policy speeds closed-loop control by 1000x","Latent policy predicts velocity-field control for new scenarios"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3625,"prompt_tokens":1079,"completion_tokens":2546,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":695,"tokens_out":2546,"duration_ms":20035,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:31:57.895641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete check: deploy the trained full-order feedback controller on a test scenario not used in training, apply each predicted control to the high-fidelity PDE, and compute the probability of arrival $P(Y(T)\\in B_{0.5}(\\mu^d_1,\\mu^d_2))$; if this probability falls well below the noise-free test levels for both the full-order loop and the latent loop, or if the latent-loop trajectory diverges from the full-order-feedback trajectory while the one-step forward errors remain at the paper's reported few-percent levels, the central claim fails.","supporting_citations":[{"cited_title":"Manzoni, A","cited_arxiv_id":null,"evidence_quote":"Supplies the constrained OCP formulation, cost functionals, and KKT/adjoint optimality system used to generate training snapshots."},{"cited_title":"Fresca, L","cited_arxiv_id":null,"evidence_quote":"Introduces the nonlinear DL-ROM autoencoder architecture whose encoder-decoder structure and joint training the approach builds on."},{"cited_title":"Hesthaven and S","cited_arxiv_id":null,"evidence_quote":"Supplies the POD-NN non-intrusive ROM idea of replacing full-order maps by neural networks trained on reduced snapshots."},{"cited_title":"Camacho and C","cited_arxiv_id":null,"evidence_quote":"Defines model predictive control, the surrogate-model closure concept that motivates the paper's latent forward model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adjoint-based optimal-control solver used to compute the full-order optimal trajectories in the data-generation phase."}],"review_version":1}