{"id":"280ab9fb-850c-496a-a6da-eabaa81fa572","arxiv_id":"2506.13369","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Optimized mutual-information sensor placement improves sparse-sensor reconstruction of cylinder-wake flows with a guided diffusion model, but the gains vanish beyond about 25 sensors.","lead":"This paper combines information-theoretic sensor placement with a physics-guided diffusion model to reconstruct flow fields from sparse measurements. On a 2D cylinder-wake test case, optimized sensor placement sharply reduces reconstruction error when only 9 to 16 sensors are available.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sensor positions are selected using the full dataset including test cases, so the reported sparse-sensor improvements may be in-sample artifacts rather than generalizable gains.","rationale":"The reader identified the same load-bearing concern: sensor placement uses full-dataset statistics, so the comparison to structured placement is potentially biased. Section IV explicitly acknowledges 'optimal fixed sensor positions based on the entirety of the dataset.' Because the POD basis and covariance feed directly into the greedy selection objective, the test cases influence which locations are chosen. The structured baseline is entirely data-independent. The percentage improvements are computed on those same test cases, so the headline numbers (83%/90% at 16 sensors, 55%/60% at 9 sensors) are exactly the quantities most vulnerable to in-sample selection. This is not a question of modeling assumptions or consensus; it is a question of whether the experimental protocol supports the generalization claim. The other weaknesses (only 16 sampling scenarios, undefined normalization of the reported L2 error, undisclosed ζ_obs, ζ_pde and the trapezoidal/gradient update ordering in Algorithm 2) are real but secondary; they would affect the absolute error values and reproducibility, not the core comparison. The proposed test—retrain the MI placement on training data only and re-evaluate on held-out cases—is the minimal experiment that would adjudicate the concern. If the improvements persist, the paper's central claim survives; if they collapse, the claim must be downgraded. Since the reader's verdict is already CONDITIONAL and this test is exactly the kind of condition needed, the verdict remains UNCHANGED.","tokens_in":13151,"tokens_out":1648,"duration_ms":17567,"concrete_test":"Re-run the sensor placement pipeline using only the 250 training simulations: compute the POD basis and covariance from those 250 cases only, run the greedy MI selection for k = 9, 16, 25, 36, 49, 64, then evaluate reconstruction L2 error on the 50 held-out test simulations (or at least the 4 test cases) with the same guided diffusion model and identical error normalization. If the optimized-vs-structured improvement at 9 and 16 sensors remains comparable (e.g., 60% or more), the concern is resolved. If the improvement drops substantially (e.g., below 20–30%) or becomes negative, the reported gains are dominated by selection bias and the central claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is that optimized MI-based placement beats structured placement by 55–90% in L2 error at 9–16 sensors. The support for this claim rests on how the sensor locations were chosen. Section IV states: 'the ROM-informed mutual information sensor placement method identifies optimal fixed sensor positions based on the entirety of the dataset, inherently averaging over different timesteps.' The POD basis Φ and covariance Σ_a used in the greedy MI objective (Eqs. 11–13) are computed from the full 300-simulation dataset, and the 50 test simulations are a subset of that dataset. Thus the sensor locations are optimized to maximize information about the very cases on which they are later evaluated. The structured baseline is generic and not adapted to any data. The comparison is therefore not between two deployment strategies evaluated out-of-sample; it is between a strategy that has 'seen' the test cases (through the POD basis and covariance) and one that has not. The reported 83%/90% improvements at 16 sensors could largely reflect this in-sample optimization advantage rather than a property of the MI placement method that would transfer to unseen flows. The paper's own explanation of occasional negative improvements—that fixed sensors cannot be optimal at every timestep—does not address this selection bias; it only explains variance within the same in-sample evaluation. Other issues (small test set, undefined error normalization, undisclosed guidance hyperparameters ζ_obs, ζ_pde and the deterministic-sampler details) compound the uncertainty, but the full-data sensor selection is the load-bearing flaw because it directly inflates the headline numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dual-guided framework for reconstructing unsteady incompressible flow fields from sparse sensor observations. First, sensor locations are selected by a greedy mutual-information maximization criterion applied to a POD-based reduced-order model (Eqs. 11-13, Algorithm 1). Second, a denoising diffusion probabilistic model (DDPM) is guided during sampling by both the observed sensor values and a divergence-free constraint from the incompressible Navier-Stokes equations (Eqs. 18-21, Algorithm 2). The method is evaluated on 2D laminar cylinder-wake flow simulations with 300 parameterized cases, comparing structured uniform sensor grids with the optimized placements at sensor counts from 9 to 64. The central empirical claim is that at 9-16 sensors, optimized placement reduces L2 reconstruction error by 55-90% relative to structured placement, while the two strategies converge for roughly 25 or more sensors.","tokens_in":13408,"tokens_out":3867,"duration_ms":40761,"significance":"If established, the proposed connection between ROM-based information-theoretic sensor placement and physics-guided generative reconstruction is a valuable contribution, potentially improving sparse-sensing reconstructions in practical flow-monitoring applications. The paper provides clear algorithmic descriptions (Algorithms 1 and 2), a reproducible experimental setup based on a public diffusion-model codebase, and a physically meaningful test case (laminar vortex shedding). However, the central quantitative claim is currently weakened by a selection-bias issue: the sensor placement uses the full dataset, including the test simulations, and the evaluation is based on only 4 test cases and 4 timesteps. The error metric is also incompletely defined, and key hyperparameters are unreported. These issues must be addressed before the claimed improvements can be considered established.","major_comments":[{"comment":"The paper states in Section IV that the ROM-informed mutual information sensor placement method 'identifies optimal fixed sensor positions based on the entirety of the dataset,' and the POD basis and covariance in Eqs. (11)-(13) are computed from all 300 simulations, which include the 50 'test' simulations. The structured baseline is generic and not informed by data. Consequently, the reported 55-90% improvements at 9-16 sensors may reflect in-sample selection bias rather than a generalizable property of MI-based placement. Please re-compute sensor placement using only the 250 training simulations and evaluate on the 50 held-out cases, or at least report both in-sample and out-of-sample improvements. This is load-bearing for the paper's claim of effectiveness in 'data-limited regimes,' since in practice sensor positions must be fixed before test data are available.","section":"Section IV (sensor placement based on entirety of dataset)"},{"comment":"The evaluation uses only 4 test cases and 4 timesteps (16 scenarios) for each sensor-count and placement strategy, and the sensor locations are optimized with access to those same cases. The aggregated L2 errors and percentage improvements in Fig. 4 therefore rest on very limited statistical evidence. Please report per-case and per-timestep error tables or plots, confidence intervals or paired significance tests, and explicitly state which 4 test cases and timesteps were used. Without this, the claim that optimized placement 'systematically' outperforms structured placement at low sensor counts is not supported.","section":"Section IV (test set size and statistical support)"},{"comment":"The L2 error is never defined. It is unclear whether the reported values are normalized by a reference velocity, by the field norm, or by the number of grid points, and whether the error is computed on the 128x128 wake grid over the full field or only in a subregion. Without a precise definition (e.g., ||v_hat - v||_2 / ||v||_2 over the wake box, averaged over cases and timesteps), the absolute values such as 0.05 versus 0.816 are uninterpretable, and the claimed 'accurate reconstructions' cannot be assessed. Please define the error metric exactly and specify the normalization.","section":"Section IV (error metric unnormalized)"},{"comment":"Key hyperparameters that directly affect the results are not reported: the observation guidance weight zeta_obs, the PDE guidance weight zeta_pde, the POD rank r, the noise variance sigma^2 used in the sensor-selection objective, and the number of sampling steps N. Since reconstruction accuracy is known to be sensitive to guidance weights, and sensor placements depend on r and sigma^2, the experiments cannot be replicated or their robustness assessed. Please report all hyperparameters in a table, including how their values were chosen and whether they were tuned separately for each sensor count.","section":"Section II and Appendix B (unreported hyperparameters)"}],"minor_comments":[{"comment":"The text introduces Eq. (4) as 'the mutual information between the full state u and the reduced representation a,' but the criterion actually optimized in Eq. (12) is I(y;a), where y=P u are sensor observations. Please clarify that the method maximizes the MI between the observed measurements and the reduced coordinates, not directly between u and a.","section":"Section II.A (notation I(u;a) vs I(y;a))"},{"comment":"The improvement percentage is defined by Eq. (25) only after Fig. 4 has already shown percentages above the bars. Please move the definition before the first use in the text.","section":"Section IV (Eq. 25 placement)"},{"comment":"The explanation of negative improvements via the fixed-sensor limitation is reasonable, but it does not address the more serious selection-bias concern raised above; please add a sentence explicitly acknowledging that the current sensor placements are computed with access to the test cases and that a training-only computation is needed to assess generalization.","section":"Section IV (negative improvements discussion)"},{"comment":"In Algorithm 2, the guidance updates are applied after the trapezoidal correction, including for the final iteration i=N-1 before returning x_N. This is consistent with returning a guided sample, but the presentation could be made more explicit if this is the intended behavior.","section":"Algorithm 2 (clarity of final update)"},{"comment":"The last row of Table II, 'Sampling Frequency (Hz)', contains the fragment '100×F= 100 samples per run -' which appears incomplete. Please reformat to state clearly that 100 snapshots per shedding cycle are extracted.","section":"Table II (formatting)"},{"comment":"The Data Availability statement says data are available 'upon reasonable request' but no code is mentioned. Given the paper's reproducibility-oriented appendices, a statement on whether the code will be released would be useful.","section":"General (code availability)"}],"recommendation":"major_revision","confidential_remarks":"The central idea is publishable in principle, but the evaluation protocol currently makes the headline quantitative claims unreliable. The strongest risk is the in-sample sensor placement: because Section IV explicitly says sensor positions are chosen from the entirety of the dataset, the comparison against a non-adaptive structured baseline is not a fair out-of-sample test. I would encourage the editor to require the authors to re-run the sensor placement on the training split only (and, if feasible, increase the number of test cases). If the qualitative advantage persists out-of-sample, the paper would be a solid contribution. The undefined L2 normalization and missing hyperparameters are also blocking issues for reproducibility. I do not recommend rejection because the methodology is sound in principle and the limitations appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the method is a reasonable integration of established components, and the qualitative pattern is clear, but the central quantitative claim is not cleanly supported because the sensor positions are optimized using the full dataset including the test cases. That is not how you measure a deployment strategy.\n\nWhat is genuinely new is the specific combination: greedy MI sensor selection on a POD manifold feeding into a DPS-style guided DDPM, with both observation and PDE guidance. Each piece exists, but the integration and the explicit comparison of structured versus optimized sensors for diffusion-based reconstruction is not something I have seen in the cited literature. The writing is clear, the pipeline is sensible, and the qualitative finding that sensor placement matters in the sparse regime and saturates beyond roughly 25 points is plausible.\n\nThe soft spot is the evaluation. Section IV states plainly that the MI placement method 'identifies optimal fixed sensor positions based on the entirety of the dataset.' The POD basis and covariance used in the greedy objective come from all 300 simulations, and the 50 test simulations are a subset. So the optimized sensor locations have effectively seen the test distribution; the structured grid has not. The 55–90% improvements at 9–16 sensors could be largely a selection-bias artifact. The paper's explanation of negative improvements, that fixed sensors cannot be optimal at every timestep, addresses variance within the test cases, not the legitimacy of the comparison.\n\nSmaller issues: only 4 test cases and 4 timesteps are actually sampled for the headline numbers; the L2 error normalization is never defined; the guidance weights, observation noise variance, and POD rank are not reported; and the diffusion sampler settings are undocumented. All are fixable.\n\nI do not think the central idea is wrong. The method should be evaluated the way it would be deployed: choose sensor positions from training data only, then test on unseen cases. If the gains survive that, the paper is a useful method-level contribution for experimental fluid dynamics. As it stands, the quantitative claims are not trustworthy, but the framework deserves a serious referee and a revision rather than a desk reject.","headline":"A sensible integration of MI-based sensor placement and guided diffusion, but the headline sparse-sensor gains are in-sample because sensor positions are chosen from the full dataset.","tokens_in":13996,"tokens_out":2378,"would_cite":false,"duration_ms":22624,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["76D05","68T07","62B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Optimal sensor placement computed from a reduced-order model lets a guided diffusion model reconstruct unsteady wake flows from as few as 9 or 16 sensors, while the advantage over uniform grids disappears beyond about 25 sensors.","keywords":["sparse flow reconstruction","sensor placement","mutual information","proper orthogonal decomposition","denoising diffusion probabilistic models","physics-informed guidance","laminar cylinder wake","reduced-order models"],"falsifier":"Select sensors using only a strict subset of the 300 simulations, freeze those coordinates, reconstruct the remaining held-out runs with the same guided diffusion model at 9 and 16 points, and compare $\\ell^2$ errors against a uniform grid; if the improvement drops substantially below the reported 55-90% in either velocity component, the gains are evidence of in-sample sensor tuning rather than of the placement criterion itself.","tokens_in":12941,"feed_emoji":"📡","tokens_out":10956,"duration_ms":103715,"temperature":0.7,"pith_summary":"The paper is trying to show that in sparse flow-field reconstruction, sensor placement should be treated as part of the generative model and can be optimized cheaply on a reduced-order representation of the flow. It selects sensor points by greedy maximization of the mutual information between a sensor's readings and the coefficients of a low-dimensional POD (proper orthogonal decomposition) basis of the velocity field. Those positions feed a denoising diffusion model that reconstructs the full field under two guides: consistency with the observed sensor values and consistency with the incompressibility (divergence-free) part of the Navier-Stokes equations. In 2D laminar cylinder-wake experiments, optimized layouts outperform uniform structured grids by 83% and 90% in $\\ell^2$ error for $v_x$ and $v_y$ at 16 sensors, and by 55% and 60% at 9 sensors. The same experiments show that the advantage mostly disappears beyond about 25 sensors, where both layouts reach errors near 0.05.","feed_headline":"With 16 flow sensors, placement cuts error by 83 to 90 percent","feed_subtitle":"Reduced-order-informed placement beats uniform grids at 9-16 sensors; both converge near 25.","key_machinery":"The load-bearing object is the Gaussian mutual-information identity $I(y;a)=\\frac{1}{2}\\log\\det(I+\\sigma^{-2}\\Phi_S\\Sigma_a\\Phi_S^\\top)$, which turns the combinatorial sensor-placement problem into a greedy selection: at each step, choose the candidate location whose addition gives the largest marginal increase in this determinant. On the generation side, the second mechanism is guided diffusion sampling: at each denoising step, a deterministic estimate $\\hat{x}_N^i$ is updated by gradients of the observation mismatch $\\|y_{\\mathrm{obs}}-M(\\hat{x}_N^i)\\|_2^2$ and of the PDE residual $\\|f(\\hat{x}_N^i)\\|_2^2$, with $f$ taken here as the divergence-free constraint $\\nabla\\cdot v=0$. Together these two mechanisms are what the paper calls dual guidance: ROM-informed sensor selection plus physics-constrained generative reconstruction.","core_discovery":"The central claim is that a greedy mutual-information objective computed from a POD-reduced representation of the flow selects sensor locations that are much better for diffusion-based reconstruction than uniform coverage, and that this advantage is concentrated in the data-scarce regime where reconstruction is otherwise unreliable. Formally, with sensor selection matrix $P$, POD basis $\\Phi$, reduced-coordinate covariance $\\Sigma_a$, and noise variance $\\sigma^2$, the mutual information between sensor readings and the reduced state is $I(y;a) = \\tfrac{1}{2}\\log\\det(I + \\sigma^{-2} P\\Phi\\Sigma_a\\Phi^\\top P^\\top)$. Maximizing this quantity greedily yields sensors concentrated near vortex cores and shear layers. When those points are used as conditioning for a denoising diffusion model, the paper reports $\\ell^2$ errors near 0.05 for $v_x$ and $v_y$ once roughly 25 points are available, with the largest relative gains at 9 and 16 points: at 16 points the improvement over a uniform grid is 83% for $v_x$ and 90% for $v_y$, and at 9 points it is 55% and 60%. The paper further claims that beyond about 25 sensors the two placement strategies converge, because the measurement set already resolves the dominant flow structures.","pith_inferences":["Beyond the paper: because the sensor positions are selected using POD statistics of the full dataset, including the test cases, part of the reported 55-90% gain may be in-sample tuning; a held-out protocol that fixes sensors using only training runs would separate placement quality from selection bias.","Beyond the paper: the authors note that fixed sensors cannot be optimal at every timestep and report negative improvements at some timesteps; a phase-adaptive or online sensor-selection policy could recover that lost margin in unsteady wakes.","Beyond the paper: the saturation near 25 sensors indicates that the effective information dimension of this laminar wake is low; for turbulent or 3D flows with higher-dimensional active manifolds, the saturation threshold should rise and the value of MI-based placement would likely persist longer."],"forward_implications":["At 9 or 16 observation points, sensor-location choice can move reconstruction from unusable (structured errors above 0.8 in some cases) to practically accurate ($\\ell^2$ error around 0.05), so placement is the deciding factor in data-limited flow reconstruction.","Beyond about 25 sensors, structured and optimized layouts converge to similar errors, meaning expensive placement optimization is unnecessary when the observation budget already resolves the dominant flow structures.","Because sensor selection runs on a POD basis and covariance rather than on full-order brute-force search, the placement step is cheap enough to apply to large spatial domains.","The dual guidance of observation consistency plus the divergence-free constraint is what allows the diffusion model to extrapolate from sparse points; the same guided diffusion model, when conditioned on uniform-grid sensors, fails only when sensors are few."],"supporting_citations":[{"why":"Supplies the underlying diffusion-based PDE solver under partial observation that this framework extends to jointly optimize sensor placement.","marker":"[1]"},{"why":"Provides the diffusion posterior sampling gradient used to condition generation on sparse observations.","marker":"[23]"},{"why":"Establishes the gappy-POD idea of choosing sensors to recover modal coefficients, which motivates the ROM-based selection criterion.","marker":"[25]"},{"why":"Demonstrates that optimized sparse sensor placement improves reconstruction and supplies the comparison context for the MI-based approach.","marker":"[28]"},{"why":"Gives the score-based ODE denoising dynamics and the score-parameterization formula on which the guided sampler is built.","marker":"[33]"},{"why":"Provides the diffusion-model design-space configuration and preconditioning used for the trained denoiser.","marker":"[39]"}],"fun_headline_variants":["Optimized sensor placement cuts sparse-flow error by up to 90%","ROM-chosen sensors beat uniform grids until about 25 points","Mutual-information sensor layout wins for sparse flow data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that sensor locations can be fixed using flow statistics drawn from the full dataset, including the cases and time steps used for evaluation; if sensors must be chosen from training data alone and transferred to unseen flow conditions, the reported advantage over structured layouts may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Optimized sensor placement cuts sparse-flow error by up to 90%","ROM-chosen sensors beat uniform grids until about 25 points","Mutual-information sensor layout wins for sparse flow data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001462,"raw_usage":{"total_tokens":5913,"prompt_tokens":1010,"completion_tokens":4903,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":4846}},"tokens_in":626,"tokens_out":4903,"duration_ms":36439,"temperature":1.0,"reasoning_tokens":4846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:03:35.154681+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select sensors using only a strict subset of the 300 simulations, freeze those coordinates, reconstruct the remaining held-out runs with the same guided diffusion model at 9 and 16 points, and compare $\\ell^2$ errors against a uniform grid; if the improvement drops substantially below the reported 55-90% in either velocity component, the gains are evidence of in-sample sensor tuning rather than of the placement criterion itself.","supporting_citations":[{"cited_title":"Gao , author S","cited_arxiv_id":null,"evidence_quote":"Establishes the gappy-POD idea of choosing sensors to recover modal coefficients, which motivates the ROM-based selection criterion."},{"cited_title":"Manohar , author B","cited_arxiv_id":null,"evidence_quote":"Gives the score-based ODE denoising dynamics and the score-parameterization formula on which the guided sampler is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the diffusion-model design-space configuration and preconditioning used for the trained denoiser."}],"review_version":2}