{"id":"4743d090-a1b9-4538-affd-f35041e08132","arxiv_id":"2502.08784","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A learned 1D wave-equation latent model with a trainable dissipation layer, driven by model predictive control, focuses or suppresses acoustic energy using sparse cylindrical scatterers in simulation.","lead":"A robot controller that moves a few small cylinders is trained to focus or suppress sound waves by learning a compressed internal model of the wave dynamics. Why read it: it is a candidate recipe for controlling waves with sparse, slow, partially observing robots, with potential uses in ultrasound devices and noise control.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The latent 1D wave model's generalization is the load-bearing risk: all MPC experiments use one fixed source location, so the learned PML and exogenous terms could be fitted to that geometry rather than to the wave equation itself.","rationale":"The reader's weakest assumption and my own load-bearing concern coincide: the learned 1D latent wave equation, including the trainable PML, must faithfully predict scalar energy for action sequences and configurations beyond the training distribution. The paper's framing as a physically realizable, interpretable method makes this external-validity question central, not peripheral. I agree with the reader that the in-distribution closed-loop results are real evidence: Table I shows AEM matching GBO on suppression and beating NODE on focusing, and the true-simulator evaluation prevents the most obvious form of model exploitation. I therefore do not recommend rejecting or changing the CONDITIONAL verdict. The missing piece is a deliberate test of whether the latent model respects symmetries of the wave equation, such as rotation of the source/scatterer configuration. A 90° rotation is a physically equivalent problem, so a physical latent model should generalize with no retraining; if it does not, the PML and exogenous terms are overfit to the training geometry. This test directly probes the 1D-sufficiency assumption and is cheap to run because the code and simulator are public. Until such a test is reported, the generalizability of the central claim remains unproven, which supports the CONDITIONAL verdict rather than full acceptance.","tokens_in":11749,"tokens_out":9181,"duration_ms":106631,"concrete_test":"Train AEM exactly as in the paper with the source fixed at (−10, 0). Then, without any retraining, run the same MPC suppression task on the true simulator with the source moved to (0, −10), a 90° rotation of the identical physical setup, and compare steady-state scattered energy against the GBO value and the NODE baseline. If the AEM-to-GBO agreement degrades beyond the 12-run error bars while GBO remains essentially unchanged, the latent model's success is tied to training-source geometry rather than to wave physics. As a control, retrain the model on the rotated source and verify performance recovers to the original level.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is MPC (Eq. 12) optimized against the learned 1D latent dynamics z, constrained to a 1D wave-equation-like NΓ with encoder-produced initial conditions and exogenous absorption l(¯x) (Eqs. 6–10). Every line of evidence that makes the method credible—long-horizon prediction in Figure 3 and the suppression results in Table I—uses a fixed source at (−10.0, 0.0). Figure 3 is shown for a single ring configuration and reports no quantitative error; Table I never changes the source. The trainable PML is a fitted sink: the only ablation (removing it) shows divergence, but no experiment separates 'physical open-space dissipation' from 'absorption tuned to the training source direction and distance.' Because l(¯x) and c(¯x,t) are learned functions on a 1D domain, there is no architectural guarantee that the projection is invariant to translation or rotation of the geometry. If the latent model has learned a scalar energy envelope for this source and scatterer family rather than the underlying wave physics, the optimal actions selected by MPC will not transfer to a rotated or displaced source, to novel scatterer arrangements, or to the real-world settings the paper explicitly aims at. This 1D-sufficiency assumption is therefore the least secure load-bearing condition: the strong in-distribution results do not refute it, because a sufficiently flexible fitted dissipation sink could reproduce those same results without capturing the true wave dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an 'Acoustic Energy Manipulation' (AEM) framework for robotic control of acoustic wave fields through sparse actuation of cylindrical scatterers. The agent observes partial wave-field images and learns a low-dimensional latent 1D wave-equation model z(x̄,t) whose dynamics are constrained to a 1D PDE with learned initial conditions, a learned exogenous term l(μ)(x̄) that includes a trainable perfectly matched layer (PML), and a learned control-dependent forcing c(φ)(x̄,t). Model predictive control (MPC) is then applied on this latent surrogate to produce piecewise-constant actions (scatterer positions/radii) for energy focusing or suppression. Experiments compare AEM against a NODE-based dynamics baseline and, for suppression, against a classical semi-analytical gradient-based optimization (GBO) solution. The paper reports that AEM outperforms NODE and random control, approximately matches GBO steady-state suppression values, and produces stable long-horizon predictions when the trainable PML is included. The authors also claim superior computational complexity relative to the learning baseline and release code and videos.","tokens_in":12096,"tokens_out":2428,"duration_ms":25817,"significance":"If the results hold, AEM would be a meaningful step toward interpretable, data-driven control of PDE-governed systems with sparse actuation and partial observability, and it is one of the few works to benchmark a learned controller against an independent semi-analytical acoustics solver. The paper's strengths include the use of a physics-constrained latent model rather than a black-box surrogate, evaluation of MPC actions on the ground-truth simulator rather than only on the learned model, benchmarking against the independent GBO method on suppression, equal-parameter-count comparison with NODE, and public code release. However, the generalization claims rest on experiments with a fixed source location and a limited set of configurations, and the computational-complexity claim is not substantiated. The 1D-sufficiency assumption and the role of the trainable PML as a fitted sink are the main risks to the central claim.","major_comments":[{"comment":"The long-horizon prediction result, which is central to the claim that the trainable PML enables stable predictions beyond the training horizon, is shown only for a single ring configuration and is presented as a visual curve without quantitative error metrics. Please report numerical errors (e.g., normalized RMSE or relative L2 error over the 200-step horizon) for AEM, NODE, and AEM-without-PML, and ideally for more than one configuration, so the reader can assess the magnitude of the divergence and the statistical significance of the improvement.","section":"Section VI.D.1, Figure 3"},{"comment":"All MPC experiments and the prediction experiment fix the acoustic source at (−10.0, 0.0). Because the trainable PML l(μ)(x̄) and the exogenous function are learned functions on the 1D domain, the demonstrated suppression and focusing results could in principle be achieved by a sink that is tuned to this particular source location and scatterer geometry rather than by a latent model that captures the underlying wave physics. To support the claimed generalization to novel configurations and real-world settings, please include at least one experiment that varies the source location or orientation (e.g., source at a different position, or the scatterer ensemble rotated relative to the source) and report whether AEM's MPC performance and prediction accuracy are maintained. Without such a test, the 1D-sufficiency assumption in Eqs. (8)–(10) remains unvalidated beyond the training distribution.","section":"Section VI.A and Table I"},{"comment":"The abstract and Section VI.D.2 claim that AEM is 'better in terms of a solution quality and computational complexity' than the learning baseline, but no runtime measurements or complexity analysis are provided. The quality comparison is supported by Table I, but the computational-complexity claim is not. Please add wall-clock training/inference times or an asymptotic complexity comparison (e.g., number of parameters, FLOPs per MPC call, or time per MPC iteration) for AEM versus NODE.","section":"Abstract and Section VI.D.2"},{"comment":"The comparison to GBO is only reported for the suppression task; for focusing, GBO is marked N/A. The conclusion that AEM is 'competitive with a classical semi-analytical method' should therefore be scoped to the suppression setting. Additionally, in the M=2 (F) suppression row both AEM and GBO report 0.38 ± 0.02; please clarify whether these are independent evaluations and, if so, provide a statistical test or at least state that the difference is within noise.","section":"Table I and Section VI.D.2"}],"minor_comments":[{"comment":"Eq. (4) uses the notation T (ζ(x,t)) = σ(t) but does not define the domain of the operator T or the norm used in the integral; please state that the norm is the L2 norm in time and specify how the scattered energy is computed from the discretized field.","section":"Section IV"},{"comment":"The wave encoder is written as Wμ : X(t) → [g(x̄,t), l(x̄)], but g appears to be an initial condition at time ti, not a function of t. Please clarify the time dependence of g and align Eq. (10) with the encoder output.","section":"Section V.A, Eq. (6)"},{"comment":"The policy is written as a(t) = π(ζ(x,t), d(t)), but the MPC formulation in Eq. (12) optimizes a sequence of piecewise-constant actions. Please clarify whether π outputs a single action or an entire sequence, and how this relates to the time-scale separation described in Section III.C.","section":"Section III.B, Eq. (2)"},{"comment":"The three actuation modes are denoted R, P, and F, but Table I only uses P and F. Please clarify whether R appears in the long-term prediction experiment only (Figure 3) and, if so, state this explicitly.","section":"Section VI.B"},{"comment":"There are several formatting and reference issues: the DOI in reference [23] is spelled out incorrectly ('10.3390/math9 222 862'), and the caption of Figure 1 uses '¶' instead of a paragraph symbol; the wording 'PDE consistency' in Section V.A would be clearer as 'PDE residual loss.'","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid proof-of-concept with a reproducible experimental setup and an honest comparison against GBO where available. The main concern is that the load-bearing generalization claim—that the learned 1D latent model with trainable PML captures the wave physics rather than the specific source/geometry configuration—is not tested. If the authors can add a source-location or geometry-variation experiment, or clearly scope the claims to the fixed-source setting, the manuscript would be much stronger. The computational-complexity claim should be either substantiated or removed from the abstract until supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper is a credible proof of concept for data-driven control of acoustic waves with sparse, slow actuation and partial observability. It deserves a serious referee, but two claims need closer scrutiny: the computational complexity advantage and the generalization of the learned latent wave model beyond a single source location.\n\nWhat's new: the synthesis. The idea of learning a latent 1D wave equation with a trainable perfectly matched layer, then using MPC on that surrogate, is not in the cited prior work. The code is public, and the suppression results are benchmarked against an independent semi-analytical GBO solution, matching it within error bars in all four configurations. The MPC actions are evaluated on the true 2D simulator, so the main evaluation isn't circular. That's the paper's strongest evidence.\n\nSoft spots. The abstract and Section VI.D claim better computational complexity than the NODE baseline, but no timings, FLOP counts, or scaling analysis appear anywhere. That claim is currently unsubstantiated. Second, every control experiment uses a fixed source at (-10,0). The trainable PML and the exogenous term l(\\bar{x}) are learned from data, so nothing rules out that they're fitted to that geometry rather than to the actual physics. The ablation shows only that removing the PML leads to divergence; it doesn't show the learned dissipation transfers to a rotated or displaced source, more scatterers, or a novel arrangement. Third, the focusing results have no independent reference (GBO is N/A), so we only know AEM beats the NODE and random baselines. That's useful, but weaker than the suppression evidence.\n\nThe math is mostly coherent: the latent dynamics are constrained to a 1D wave equation and integrated numerically, so the surrogate is a valid wave solution in latent space. The 'interpretability' claim is oversold, though—the encoders are CNNs, and the physical meaning of g, c, l is assigned after the fact rather than enforced.\n\nWho it's for: researchers in data-driven PDE control, robotics with sparse actuation, and acoustic metamaterials. Not a real-world system paper, but a solid simulation-level proof of concept.\n\nRecommendation: send it to peer review. Ask the authors to provide actual complexity measurements or temper the claim, and to add at least one experiment with a rotated or shifted source. If the PML transfer holds, the paper is strong; if not, the generalization claim reduces to in-distribution curve fitting.\n\nBest","headline":"Plausible new synthesis for acoustic wave control; suppression results are solid, but the complexity claim is unmeasured and the fixed-source evaluation leaves the generalization claim unproven.","tokens_in":12643,"tokens_out":3906,"would_cite":false,"duration_ms":36302,"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":"The paper claims a robot can steer acoustic waves by learning a 1D wave-equation surrogate of a partially observed 2D wave field and planning sparse scatterer actions with model predictive control.","keywords":["acoustic wave manipulation","model predictive control","latent wave equation","perfectly matched layer","sparse robotic actuation","partial differential equations","physics-informed learning"],"falsifier":"Run the trained AEM on a configuration it never saw, for example eight scatterers or a source location different from $(-10.0, 0.0)$, and compare its long-horizon predicted scattered energy against the ground-truth PDE simulation; if the prediction diverges substantially while the NODE baseline tracks better, or if the MPC actions fail to suppress or focus energy in the true environment, the central claim of a transferable 1D latent wave model is falsified.","tokens_in":11529,"feed_emoji":"🌊","tokens_out":4708,"duration_ms":42457,"temperature":0.7,"pith_summary":"The paper tries to establish that a robot with only partial sensor images of an acoustic wave field can learn to manipulate that field through sparse scatterer actuation by compressing the wave into a 1D latent wave equation. The method couples two neural encoders with numerical integration of the latent PDE and a trainable perfectly matched layer for open-space dissipation. A sympathetic reading of the experiments is that this physics-constrained surrogate predicts scattered energy more faithfully over long horizons than a neural-ODE baseline and supports model-predictive control that both focuses and suppresses energy. If true, the result matters because it offers an interpretable and computationally cheaper route to controlling PDE-governed environments, not just acoustics, without full state observability.","feed_headline":"Learned 1D wave model steers sound with sparse robots","feed_subtitle":"A physics-constrained surrogate plans scatterer actions to focus or silence acoustic energy, and matches a classical optimizer.","key_machinery":"The load-bearing object is the learned latent 1D wave equation $\\partial_t z(\\bar{x},t) = \\mathcal{N}^{\\Gamma}(z; c_\\phi(\\bar{x},t), l_\\mu(\\bar{x}), t)$, with $c_\\phi$ coming from the robot encoder $D_\\phi$ and $l_\\mu$ from the wave encoder $W_\\mu$. The wave encoder maps a sensor image to an initial condition $g_\\mu(\\bar{x},t_i)$ and an exogenous function $l_\\mu(\\bar{x})$ that encodes energy excitation and the trainable perfectly matched layer, a fitted damping sink that absorbs outgoing waves; the robot encoder maps the known robot trajectory to the latent control term $c_\\phi$. The latent trajectory is produced by numerical integration, which guarantees it satisfies the chosen PDE and introduces no trainable dynamics network, and the training loss combines energy prediction error, PDE consistency, and initial-condition error. This structure is what allows model predictive control to optimize actions through the surrogate.","core_discovery":"The central discovery is that a full 2D acoustic wave field, partially observed through sensor images and controlled through sparse cylindrical scatterers, can be compressed into a 1D latent wave equation whose evolution is obtained by numerical integration rather than learned as a black box. With a trainable perfectly matched layer included as an exogenous function $l_\\mu(\\bar{x})$ in the latent space, this model predicts scattered energy over horizons far longer than training and serves as the dynamics model for model predictive control. In experiments, the resulting controller focuses more energy and suppresses more scattered energy than a neural-ODE baseline with equal parameter count, and its steady-state suppression matches a classical gradient-based optimization solution where one is known.","pith_inferences":["A natural next test, not in the paper, is whether the trainable PML learned in one configuration transfers to a different domain size or incident frequency; the paper shows only that removing it causes divergence, not that the fitted dissipation generalizes.","Because the latent model is a 1D wave equation, the same architecture could in principle control other hyperbolic PDEs with a scalar energy target, such as elastic or electromagnetic wave fields; the paper claims generality but tests only acoustics.","The equal-parameter comparison with NODE suggests the gain comes from the wave-equation inductive bias rather than model capacity; a clean ablation would swap the latent PDE for an ODE or a different PDE family to isolate that bias."],"forward_implications":["Long-horizon prediction of scattered energy stays accurate far beyond the 20-step training horizon when the latent model includes a trainable perfectly matched layer.","MPC using the latent 1D model focuses more energy and suppresses more scattered energy than neural-ODE MPC across positional configurations with $M=1,2,4$ scatterers and the fully adjustable $M=2$ configuration.","Steady-state suppression matches the semi-analytical gradient-based optimization solution, including the $M=2$ full-adjustment case where both reach a suppressed energy of 0.38.","The method needs only sensor images and known robot dynamics, not full wave-state observability, and it avoids an extra neural network for latent dynamics by using numerical integration of the 1D wave equation.","The approach is presented as a proof of concept for physically realizable wave manipulation, with the same latent-wave-equation idea suggested as applicable to other PDE-governed systems."],"supporting_citations":[{"why":"Supplies the gradient of total multiple scattering cross-section and the semi-analytical gradient-based optimization solution that serves as the classical benchmark.","marker":"[8]"},{"why":"Defines the neural ordinary differential equation baseline whose black-box latent dynamics the proposed method replaces.","marker":"[29]"},{"why":"Introduces the perfectly matched layer concept that the trainable PML in the latent space is designed to emulate.","marker":"[31]"},{"why":"Represents the earlier LSTM surrogate-dynamics approach for PDE control that motivates the partial-observability and sample-efficiency concerns.","marker":"[26]"},{"why":"Demonstrates realizable MPC with sparse cylinder control and limited sensor data for Navier-Stokes, providing the realizability choices adopted here.","marker":"[27]"},{"why":"Shows physics-informed neural networks for MPC of ODE systems, the closest prior physics-informed control work that this paper extends to PDEs.","marker":"[37]"},{"why":"Provides the physics-informed neural network framework supporting the claim that physics knowledge improves generalization and sample efficiency.","marker":"[39]"}],"fun_headline_variants":["Sparse robots steer sound with a 1D learned wave surrogate","Learned 1D wave model lets sparse robots focus or mute sound","Acoustic wave control via sparse robots and a 1D latent PDE","Sparse robots learn to focus or silence sound via 1D wave equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest premise is that a single 1D wave equation, whose initial conditions and exogenous source/dissipation term are read off by neural encoders, can faithfully predict the scalar energy signal of the true 2D wave field for unseen actions and long horizons; if the 1D model stops tracking the true energy, the MPC actions it recommends will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Sparse robots steer sound with a 1D learned wave surrogate","Learned 1D wave model lets sparse robots focus or mute sound","Acoustic wave control via sparse robots and a 1D latent PDE","Sparse robots learn to focus or silence sound via 1D wave equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3453,"prompt_tokens":885,"completion_tokens":2568,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2489}},"tokens_in":501,"tokens_out":2568,"duration_ms":15776,"temperature":1.0,"reasoning_tokens":2489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T23:41:27.522328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained AEM on a configuration it never saw, for example eight scatterers or a source location different from $(-10.0, 0.0)$, and compare its long-horizon predicted scattered energy against the ground-truth PDE simulation; if the prediction diverges substantially while the NODE baseline tracks better, or if the MPC actions fail to suppress or focus energy in the true environment, the central claim of a transferable 1D latent wave model is falsified.","supporting_citations":[{"cited_title":"The gradient of total multiple scattering cross-section and its application to acoustic cloaking,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient of total multiple scattering cross-section and the semi-analytical gradient-based optimization solution that serves as the classical benchmark."},{"cited_title":"Neu- ral ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Defines the neural ordinary differential equation baseline whose black-box latent dynamics the proposed method replaces."},{"cited_title":"A perfectly matched layer for the absorption of electromagnetic waves,","cited_arxiv_id":null,"evidence_quote":"Introduces the perfectly matched layer concept that the trainable PML in the latent space is designed to emulate."},{"cited_title":"Physics-informed neural networks-based model predictive control for multi-link manipulators,","cited_arxiv_id":null,"evidence_quote":"Shows physics-informed neural networks for MPC of ODE systems, the closest prior physics-informed control work that this paper extends to PDEs."}],"review_version":1}