{"id":"011b0c52-f4b7-440d-a7af-499a06a6f86e","arxiv_id":"2507.02500","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Bayesian C-optimal sensor placement workflow with low-rank acceleration is applied to contaminant source identification and sensor steering in complex geometries.","lead":"Researchers built a computer framework that decides where to place fixed and moving sensors so that uncertainty about a danger zone, such as where a contaminant cloud came from, is minimized. It combines Bayesian source inference with fast low-rank approximations and tests the workflow on simulated airborne contamination scenarios, though not yet on real data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-rank Hessian inverse in Eq. (9) has an internal sign error (should be (I + Hmisfit)^-1 ≈ I - V D V^T), contradicting Section 4.1; if actually implemented, the C-optimal objective is not the posterior variance and the reported sparse layouts are unsupported.","rationale":"The reader's identified weakest assumption (self-consistent synthetic validation) is a genuine limitation, but it applies equally to most OED papers and does not reveal a defect in the mathematics. The Eq. (9) inconsistency is more load-bearing because it concerns the correctness of the very quantity being optimized. If the low-rank covariance formula is wrong, the sparse layouts are not necessarily near-optimal, and the comparisons with the 96-sensor configuration are uninterpretable. I therefore focus the attack on Eq. (9). The concern is made in good faith: the paper may have a typo, as Section 4.1 contains the correct minus sign, but the manuscript as printed is internally inconsistent. The recommended verdict remains CONDITIONAL, with the additional explicit condition that Eq. (9) be corrected and the numerical experiments be re-verified with the corrected formula. Agreement with the reader is partial: they noted the sign inconsistency in the rationale but did not make it the weakest assumption.","tokens_in":13469,"tokens_out":5886,"duration_ms":61363,"concrete_test":"Independently verify Eq. (9): start from H = I + Hmisfit and apply the Sherman-Morrison-Woodbury identity to the rank-r approximation Hmisfit ≈ V_r diag(λ) V_r^T, confirming that the correct update is I - V_r diag(λ/(1+λ)) V_r^T. Then, for the OED 1 setup, evaluate the C-optimal objective c^T H^-1 c for the reported five-sensor layout using both the printed sign (I + V_r D_r V_r^T) and the corrected sign (I - V_r D_r V_r^T). If the printed sign yields a negative value for any candidate weight vector, or if the two objective values differ by more than round-off, Eq. (9) as printed is invalid; re-run OED 1 with the corrected formula and check whether the five-sensor layout and the claimed accuracy comparison survive.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing concern is an internal sign inconsistency in the low-rank approximation of the posterior covariance, Eq. (9). The paper defines the preconditioned Hessian as H_h(w) = Hmisfit_h(w) + I (from Eq. (8)), so the Woodbury identity gives (H_h)^-1 = (I + Hmisfit)^-1 ≈ I - V_r D_r V_r^T with D_r = diag(λ/(1+λ)). Eq. (9) instead states (Hmisfit_h(w) - I)^-1 ≈ I + V_r D_r V_r^T, changing both the sign inside the inverse and the sign of the correction. This is not harmless: Section 4.1 uses the corrected form, c_h^T A_h^-1 (I - V_r D_r V_r^T) A_h^-1 c_h, to evaluate the C-optimal objective, so the printed derivation and the implemented formula are mutually contradictory. If the printed Eq. (9) were actually used, the 'posterior variance' being minimized would be an indefinite quadratic form rather than a covariance, and the reported five-sensor and eight-sensor layouts in OED 1 and OED 2 would not be C-optimal in the stated sense. Since no code is provided, the reader cannot tell whether the discrepancy is a typographical slip in Eq. (9) or a genuine error in the computation; either way, the central numerical claims are not reproducible from the manuscript as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian goal-oriented optimal experimental design framework for stationary and mobile sensor placement in PDE-constrained inverse problems, applied to airborne contaminant source identification. The authors use a Laplacian-like trace-class prior, precondition the Hessian, approximate the preconditioned misfit Hessian by a rank-r eigendecomposition, and combine a C-optimality criterion with an ℓ1 penalty to select sparse sensor configurations. For dynamic steering, the QoI is centered at the maximum of the current maximum-a-posteriori reconstruction and the mobile sensor moves to the highest-weight candidate location. Three numerical experiments on a campus geometry obtained from OpenStreetMap are presented: reconstructing the initial condition in a region with five stationary sensors, monitoring a space-time QoI with eight sensors, and steering a mobile sensor alongside one fixed sensor. The paper claims accuracy comparable to a dense 96-sensor configuration and faster reconstruction with the mobile sensor.","tokens_in":13854,"tokens_out":8918,"duration_ms":101909,"significance":"If the results hold, the framework would be a useful practical extension of earlier dynamic-steering work to goal-oriented designs and complex geometries, with potential real-time capability due to the low-rank offline/online decomposition. The paper gives an explicit adjoint-based formulation for space-time quantities of interest and reports a substantial speedup for reduced-order forward evaluations. Strengths include the clear finite-element setting, the use of a trace-class prior, and the three demonstration cases. However, the validation is entirely synthetic and largely qualitative, no code or data is provided, and the printed low-rank and derivative formulas contain sign inconsistencies. The contribution is therefore promising but not yet fully supported as written.","major_comments":[{"comment":"Equation (9) is inconsistent with Equation (8) and with the formula used in Section 4.1. Since the preconditioned Hessian is Hmisfit + I (Eq. (8)), the Sherman-Morrison-Woodbury identity gives (I + Hmisfit)^{-1} ≈ I - V_r D_r V_r^T with D_r = diag(λ/(1+λ)). Equation (9) instead states (Hmisfit_h(w) - I)^{-1} ≈ I + V_r D_r V_r^T, changing both the sign inside the inverse and the sign of the low-rank correction. Section 4.1 uses c_h^T A_h^{-1}(I - V_r D_r V_r^T)A_h^{-1}c_h, so the printed derivation and the implemented objective are mutually contradictory. Because no code is provided, the reader cannot determine which form was actually used; if Eq. (9) were used, the minimized quantity would not be the posterior variance and the five- and eight-sensor layouts in Section 5 would not be C-optimal in the stated sense. Please correct the equation and provide a reproducible implementation or a numerical check of the objective.","section":"§3.2, Eq. (9)"},{"comment":"The derivative formula has the wrong sign. For Γpost(w) = (F^T W Γ_noise^{-1} W F + Γ_pr^{-1})^{-1} and the trace objective c^T H^{-1}c, the derivative with respect to w_i is -(F_i H^{-1}c)^2/σ^2, where F_i is the i-th row of the observation operator. Equation (12) states a positive value. Combined with the ℓ1 penalty of Eq. (10), a positive gradient would drive all weights to zero during minimization, so either the sign is a typographical error or the optimization used a different objective; in either case the reported sparse configurations are not supported by the printed formulas.","section":"§4.1, Eq. (12)"},{"comment":"The central claim that five (resp. eight) sensors achieve accuracy comparable to the full 96-sensor configuration is supported only by visual comparison of reconstructed fields (Figures 9 and 10) and qualitative variance plots. No error norms, signal-to-noise metrics, or posterior-variance reductions are reported. Please provide quantitative errors, e.g., relative L² errors in the quantity of interest and in the reconstructed initial condition, for both the sparse and the full configurations.","section":"§5.1 and §5.2"},{"comment":"The numerical validation is self-consistent rather than independent: the ground truth is generated by the same discretized advection-diffusion model used in the likelihood, the wind field is assumed known and divergence-free, and the noise variance and prior hyperparameters are hand-chosen. The experiments therefore do not test the method under model error, unknown wind, or sensor faults. The conclusion in Section 6 that the method proves real-world capability is too strong; please add robustness experiments (e.g., perturbed wind, different noise levels, or model discrepancy) or soften the claims accordingly.","section":"§5"},{"comment":"The rank r=200 truncation and the binarization threshold are not justified. Figure 8 shows singular-value decay for F and F∘A, but not for the preconditioned Hessian used in Eq. (9), and no convergence study of the C-optimal objective with respect to r is given. Since the design depends on the low-rank approximation, please report the spectrum of the preconditioned misfit Hessian and the sensitivity of the selected layouts to r, α, and the threshold.","section":"§3.3 and §5"}],"minor_comments":[{"comment":"In the formula for mmap(w), Γ_pr mpr should be Γ_pr^{-1} mpr; as printed, the equation is dimensionally inconsistent and omits the inverse of the prior covariance.","section":"§2.3"},{"comment":"The symbol P is overloaded as both the QoI operator and the spatial subset (e.g., P1 and P2 in Section 5); please use distinct notation for the operator and the subset.","section":"§4.1 and §5"},{"comment":"The right panel claims that the reduced-order model approximates the forward operator fairly well, but no error metric or maximum relative error is given; please add a quantitative error plot.","section":"Figure 8"},{"comment":"The mobile sensor speed of approximately 40 m/s greatly exceeds typical UAV speeds; if this is a deliberate abstraction, state this and discuss the implications for the steering claim.","section":"§5.3"},{"comment":"There are minor typographical errors, including 'Navier-Stockes' for 'Navier-Stokes', 'rigth' for 'right' in the Figure 9 caption, and inconsistent capitalization of section references.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper provides no code or data, which is a serious obstacle given the sign inconsistencies in Eq. (9) and Eq. (12). I would encourage the editor to request a reproducible implementation or at least a detailed numerical verification of the objective and its gradient. The title and abstract emphasize crisis management, but the validation is entirely synthetic; please consider whether the journal's scope requires either more realistic testing or more cautious claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a capable, clearly written assembly of existing Bayesian optimal experimental design machinery applied to airborne contaminant tracking in a realistic campus geometry. It is not a methodological breakthrough—the C-optimal goal-oriented formulation, the low-rank Hessian approximation, and the dynamic steering loop all come from the cited literature—but it is a useful integration, and the complex-geometry demonstrations are a real step beyond the earlier toy cases.\n\nThe paper does several things well. It spells out the whole pipeline—forward model, Bayesian inverse problem, adjoint-based reduction of a space-time quantity of interest to a single vector, sparse OED via ℓ1 regularization, and offline/online splitting—and it tests the pieces on three scenarios that map directly to crisis-response questions. The reduction of the time-dependent QoI to one adjoint solve (Eq. 11) is standard but cleanly presented. The reported 160x speedup from the low-rank reduced-order model is believable and worth having in the literature. The citation pattern is honest: Attia et al., Alexanderian et al., Wu et al., and Wogrin et al. are all cited and are the right background.\n\nThe soft spots are real but addressable. First, Eq. (9) has a genuine sign inconsistency: for a preconditioned Hessian of the form H_misfit + I, the Woodbury identity gives (I + H_misfit)^{-1} ≈ I - V D V^T, not I + V D V^T, and the printed (H_misfit - I)^{-1} does not match Section 4.1, which uses the correct form. Since Section 4.1 uses the correct expression, the most likely explanation is a typo, but as written the manuscript is internally contradictory and the numerical results are not reproducible from the paper. No code or data is released, so a reader cannot tell which form was actually implemented.\n\nSecond, the validation is self-consistent rather than independent: synthetic observations are generated with the same discretized forward model used for inversion and design. That is a common first step, but the paper presents the sensor-count claims as if they were established effectiveness. Unknown wind, model error, or sensor faults are exactly where a crisis-management method would live or die. The steering loop also defines the QoI around the maximum of the reconstructed initial condition, which ties the target to the current reconstruction—reasonable as a heuristic, but it weakens the word \"optimal.\" Quantitative error metrics, such as relative L2 error in the QoI region comparing five sensors against 96, are missing; the paper relies on visual comparison.\n\nWho is this for? Practitioners in applied inverse problems or crisis management who want a concrete recipe for static and mobile sensor placement. Method developers will not find new theory. It deserves a serious referee: the application is timely, the math is mostly standard and sound, and the flaws are fixable. My recommendation: send it to peer review, but require correction of Eq. (9), an explicit statement of which formula was implemented, quantitative comparisons against the full sensor configuration, and either code/data release or at least one sensitivity test against model or geometry perturbation.","headline":"A competent, clearly written integration of established Bayesian OED tools for contaminant tracking; the realistic geometry is a plus, but the method is not new and the validation is self-consistent, with a genuine sign inconsistency in Eq. (9) that must be fixed before the headline claims are fully supported.","tokens_in":14381,"tokens_out":3706,"would_cite":false,"duration_ms":44379,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N21","62K05","35R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Five goal-oriented sensors match the reconstruction accuracy of a 96-sensor grid, and a steered mobile sensor beats a stationary one within 7 seconds.","keywords":["goal-oriented optimal experimental design","C-optimal design","Bayesian inverse problems","sensor placement","dynamic sensor steering","airborne contaminant transport","low-rank Hessian approximation","advection-diffusion"],"falsifier":"Generate synthetic observations from a finer or otherwise different discretization of Equation 1, or perturb the wind field, and rerun the C-optimal five-sensor design; if reconstruction error in the region of interest no longer matches the 96-sensor configuration, the result hinges on the test setup. A controlled tracer-release experiment with independent ground truth would also settle it.","tokens_in":13226,"feed_emoji":"📡","tokens_out":6058,"duration_ms":65121,"temperature":0.7,"pith_summary":"This paper is trying to show that you do not need many sensors to locate and track an airborne contaminant source if you choose where to measure based on the specific decision you care about. It formulates source identification as a Bayesian inverse problem governed by an advection-diffusion PDE and then selects sensor locations by a C-optimal design criterion that minimizes posterior variance of a chosen quantity of interest, using a low-rank Hessian approximation to keep the computation fast. In the paper's experiments, five selected stationary sensors match the reconstruction quality of a 96-sensor grid in a critical region, eight sensors are enough to monitor a region over a time window, and one mobile sensor steered by the same criterion reaches good reconstruction in about 7 seconds. The authors claim the framework is transferable to other crisis-management sensing tasks because the criterion only depends on the model through the chosen quantity-of-interest vector.","feed_headline":"Five sensors match a 96-sensor grid for source finding","feed_subtitle":"A goal-oriented Bayesian design steers few stationary or mobile sensors to fast, reliable contaminant-source estimates.","key_machinery":"The load-bearing object is the C-optimal design criterion applied to the posterior covariance of the quantity of interest: instead of minimizing the trace of the full posterior covariance (A-optimality), one minimizes the variance of a single linear functional $\\rho = P(m)$ of the parameter. The goal-oriented operator $P$ is encoded as a vector $c$ in the finite-element space — for a space-time region of interest, $c$ is the solution of an adjoint advection-diffusion problem $c = K^*(1_{[T_0,T_{\\mathrm{final}}]\\times P})$ — and the design objective becomes $c^T H(w)^{-1} c$. This is evaluated cheaply through a low-rank approximation of the preconditioned Hessian (randomized SVD of the preconditioned forward operator plus Sherman-Morrison-Woodbury), and sparse layouts are obtained by adding an $\\ell^1$ penalty on sensor weights and solving with L-BFGS-B. The gradient of the objective with respect to each sensor weight reduces to a squared forward response $(F^i(q))^2$, which makes the optimization tractable.","core_discovery":"The paper's central claim is that goal-oriented C-optimal experimental design, accelerated by low-rank approximation of the preconditioned Hessian, produces sparse stationary sensor layouts and effective dynamic steering trajectories for PDE-constrained Bayesian inverse problems. In the airborne contaminant test case, five optimally placed stationary sensors reconstruct the initial condition in a critical subregion with accuracy comparable to the full 96-sensor configuration, eight sensors suffice for monitoring a region over a time window, and a mobile sensor steered by the same criterion substantially improves the reconstruction within about 7 seconds where a stationary sensor alone fails to give a reliable estimate within 12 seconds. The authors present this as evidence that decision-relevant uncertainty, rather than global field uncertainty, is the right target for sensor placement in crisis management.","pith_inferences":["If the same comparison were repeated with data from an independent forward model — different discretization, unknown or unsteady wind, or sensor faults — the reported sensor counts and times could shift; the paper does not establish robustness to such model error.","The five-sensor advantage is demonstrated on synthetic data in which the ground truth and the inversion share a discretization; a laboratory tracer release with measured wind would separate algorithmic merit from self-consistency.","The dynamic steering loop recentres the quantity of interest on the current maximum of the reconstruction, which is a heuristic; one could extend it to account for posterior uncertainty in the location of the maximum, or to multiple mobile sensors.","Because the goal-oriented criterion reduces to a fixed vector, the method's computational cost is dominated by the low-rank Hessian approximation, suggesting it could scale to three-dimensional crisis geometries if the spectrum of the preconditioned Hessian decays similarly."],"forward_implications":["In the static source-identification case, five sensors selected by the goal-oriented C-optimal criterion reconstruct the initial condition inside the region of interest with accuracy comparable to the full 96-sensor grid.","When the goal is monitoring a region over a time window, eight sensors suffice to reconstruct the concentration evolution and concentrate the uncertainty reduction in that region.","Steering one mobile sensor with the same design criterion, alongside a poorly placed stationary sensor, yields substantially better reconstruction after about 7 seconds than the stationary sensor alone does after 12 seconds.","Because the design objective depends on the quantity of interest through a single vector $c$, the same algorithm transfers to other physical models and other user-defined quantities of interest."],"supporting_citations":[{"why":"Supplies the dynamic sensor steering framework that this paper extends to a more complex geometry.","marker":"[1]"},{"why":"Provides the Bayesian inverse problem formulation and software approach used for prior, posterior, and Hessian actions.","marker":"[2]"},{"why":"Supplies the randomized algorithm used to compute the low-rank factorizations of forward and preconditioned operators.","marker":"[4]"},{"why":"Provides the preconditioned inexact Newton-CG solver used for the online inverse problem solution.","marker":"[6]"},{"why":"Establishes the optimal design setting for infinite-dimensional Bayesian linear inverse problems that the goal-oriented criterion builds on.","marker":"[7]"},{"why":"Develops goal-oriented optimal approximations of Bayesian linear inverse problems, the conceptual basis for minimizing uncertainty in the quantity of interest.","marker":"[8]"},{"why":"Formulates goal-oriented optimal design for large-scale Bayesian linear inverse problems, directly motivating the C-optimal objective used here.","marker":"[9]"},{"why":"Introduces the sparsified optimal design approach and the derivative formula that the paper adapts for the C-optimal objective.","marker":"[15]"},{"why":"Supplies the scalable low-rank Hessian approximation and uncertainty propagation methodology used for the posterior covariance.","marker":"[18]"}],"fun_headline_variants":["Five sensors match a 96-sensor grid for contaminant tracking","Goal-oriented Bayesian design: 5 sensors do the work of 96","Mobile sensor outperforms stationary in crisis source mapping","Low-rank approximation cuts sensor needs 19-fold for crises"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the numerical validation is meaningful despite being self-consistent: the same discretized advection-diffusion model generates the 'true' measurements, defines the likelihood, and supplies the forward map, so the claimed five-sensor accuracy may not survive model error or unknown conditions.","fun_headline_variants_meta":{"raw":{"variants":["Five sensors match a 96-sensor grid for contaminant tracking","Goal-oriented Bayesian design: 5 sensors do the work of 96","Mobile sensor outperforms stationary in crisis source mapping","Low-rank approximation cuts sensor needs 19-fold for crises"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001023,"raw_usage":{"total_tokens":4224,"prompt_tokens":767,"completion_tokens":3457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":383,"completion_tokens_details":{"reasoning_tokens":3387}},"tokens_in":383,"tokens_out":3457,"duration_ms":33131,"temperature":1.0,"reasoning_tokens":3387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:28:42.283071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic observations from a finer or otherwise different discretization of Equation 1, or perturb the wind field, and rerun the C-optimal five-sensor design; if reconstruction error in the region of interest no longer matches the 96-sensor configuration, the result hinges on the test setup. A controlled tracer-release experiment with independent ground truth would also settle it.","supporting_citations":[{"cited_title":"Goal-oriented optimal approximations of Bayesian linear inverse problems","cited_arxiv_id":null,"evidence_quote":"Develops goal-oriented optimal approximations of Bayesian linear inverse problems, the conceptual basis for minimizing uncertainty in the quantity of interest."}],"review_version":1}