{"id":"52095bc4-2583-4f02-aec9-5eaa9cf1ced4","arxiv_id":"2412.19387","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A reduced-order-model based least-squares estimator, paired with a greedy sensor placement algorithm, reconstructs freezing food temperature fields from sparse airflow measurements, but is only tested against the simulator that generated the data.","lead":"Using a data assimilation pipeline built on a reduced-order model of turbulent airflow and phase change, this paper estimates the full temperature field of a freezing salmon slice from a small number of airflow sensors, and a greedy algorithm picks where those sensors should go. A generalist reader would see a practical blueprint for real-time, non-invasive temperature monitoring in food freezing, though the validation is currently synthetic.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Forward URANS solver is validated only at Ra≈1e7–1e9 while the claimed operating regime is Ra>1e10, so the reported ~1% errors measure consistency with a possibly biased model, not physical accuracy.","rationale":"Good faith reading: the paper's central contribution is a ROM-regularized inverse pipeline plus a sensor-placement heuristic, and it is explicitly a proof-of-concept. The numerical machinery is internally consistent: POD modes from 48 training runs, greedy selected sensors, and held-out tests from 16 runs all support the claim that, conditional on the forward model, the inverse problem is well-posed and accurate. I do not see an internal inconsistency in the least-squares/POD/Greedy derivation, and the comparison against regular sensor grids gives genuine evidence that the greedy placement helps. The most load-bearing condition for the advertised real-world capability is not the linear algebra but the validity of the forward model in the regime where the method is claimed to work. That condition is unverified: the benchmark validations bracket Ra 6.81e7–1.58e9, the target regime is Ra>1e10, and the 1% errors are computed against the same solver's synthetic truth. The one experimental test is at low Ra and does not use external-only sensors. This is a gap in evidence, not a contradiction; it warrants the conditional verdict. A secondary issue is that Algorithm 1's greedy sequential maximization does not by itself prove global optimality of the selected sensor set, but the paper's empirical comparisons with regular placements do not depend on that overclaim, so it is not the load-bearing issue. Verdict unchanged.","tokens_in":28177,"tokens_out":5885,"duration_ms":58168,"concrete_test":"Run an end-to-end validation in the target regime: in a freezer or matched wind-tunnel geometry with Ra≥10^10, place thermocouples inside a salmon surrogate or food simulant, observe only airflow-domain temperatures (thermal-camera pixels or point sensors), apply the greedy ROM reconstruction of Section 5.3, and compare the inferred internal temperatures with the held-out thermocouples over the full freezing curve. If the mismatch exceeds the sub-2% level reported in Fig. 13 or scales with the forward-solver bias measured in the same rig, then the central extrapolation claim is not established; if it matches, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.4 defines the operational regime as Ra>10^10 and selects a k-ω SST URANS closure. The solver validation in Section 3.2 is against a food-freezing experiment at Ra=6.81e7 (P1) and an air-cavity convection experiment at Ra=1.58e9 (P2). Sections 5.1 and 5.3 compute all reconstruction errors relative to the same FVM solver's output, so the reported 1% (and local <2%) errors upper-bound the ROM/inverse error only under the assumption that the forward model is the truth. Section 5.2 does use experimental thermocouple data, but it is at the low Ra regime and the ROM is built from a single simulation, so it does not validate either the high-Rayleigh turbulence closure or the external-only extrapolation setting. If the URANS model carries a bias at Ra>10^10 (for example in the forced-plus-natural convection balance, shear-layer reattachment, or wall heat transfer), the estimated internal temperatures inherit that bias because the reconstruction is a projection onto states generated by the same model. The paper's own text labels the main experiments as a proof-of-concept and future work explicitly lists LES 'to strengthen the credibility of the model,' which corroborates that the central realistic-turbulence claim is not yet demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a computational framework for estimating the full temperature field of a freezing salmon slab in a ventilated freezer from a limited number of temperature measurements. The forward model couples URANS airflow with a k-omega SST turbulence model, Boussinesq buoyancy, and an effective-heat-capacity phase-change model, solved by a WENO3 finite-volume scheme. A reduced basis is constructed by truncated SVD of snapshots from 48 training simulations sampled over a five-dimensional parameter range, and the inverse step is a ROM-regularized least-squares projection of the measurements onto that basis. Sensor positions are selected by a greedy algorithm that aims to maximize the smallest singular value of the cross-Gramian W^T Phi. Numerical experiments cover full-domain thermocamera-like measurements, measurements restricted to the airflow region outside the food, and comparisons of greedy versus regularly spaced sensor layouts, together with a benchmark using experimental thermocouple data.","tokens_in":28455,"tokens_out":6956,"duration_ms":62908,"significance":"If the claimed performance held at the operational conditions, the framework would be a practically useful tool for non-invasive, real-time monitoring of food freezing. The paper has several genuine strengths: the offline/online cost split is clearly quantified in Table 3, the sensor-placement criterion is based on an a-priori bound rather than on the measured values themselves, and the authors are explicit that the main validation is a proof-of-concept based on synthetic measurements. The methodological core (POD basis plus least-squares reconstruction) is standard, and the numerical behavior is internally consistent. However, the title-level claim of optimally located sensors and the abstract's claim of extrapolation under realistic turbulent flow conditions rest on validation that is almost entirely self-consistency with the training solver; this restricts the confidence that can be placed in the physical accuracy of the estimated fields.","major_comments":[{"comment":"The derivation of the a-priori bound contains a notation error that makes the displayed algebra invalid. After writing the singular value decomposition G = U_hat S_hat V_hat^T, the denominator is printed as d^T V^T Phi^T Phi V d, using the V from the snapshot SVD instead of V_hat; as written, the replacement of this denominator by d^T d is not justified. If V_hat was intended, the equality follows only after using both the orthonormality of the columns of Phi and the orthogonality of V_hat. Please correct this step or, more directly, invoke the standard fact that, for orthonormal Phi, inf_c ||G c|| / ||Phi c|| equals the smallest singular value of G. This is load-bearing because e(n) in Eq. (14) and the greedy criterion in Algorithm 1 are both derived from this bound.","section":"Section 4.3, Eq. (13)"},{"comment":"The operating regime is defined by Ra > 10^10, and the abstract claims efficient extrapolation from external measurements under realistic turbulent flow conditions. However, the forward solver is validated in Section 3.2 at Ra = 6.81e7 (P1) and Ra = 1.58e9 (P2), and all reconstruction errors in Sections 5.1, 5.3, and 5.4 are computed against the output of the same FVM solver that generated the training snapshots. The only experimental comparison, Section 5.2, uses a single-parameter ROM built from one low-Rayleigh simulation and does not exercise the external-only, high-Rayleigh setting that is the central claim. The reported ~1% errors therefore bound the ROM and inverse error conditional on the forward model being exact, not the error against real freezing behavior at Ra > 10^10. The claims should be reframed accordingly, or the paper should include a high-Rayleigh experimental or independent benchmark.","section":"Sections 2.4, 3.2, 5.1, 5.3"},{"comment":"The statement that the greedy selection guarantees beta(m*) >= beta(m_sub) for any other selection of m measurements is stronger than what a stepwise greedy procedure can establish. Greedy algorithms of this type are locally optimal at each addition but are not globally optimal over all m-sensor sets in general. This overstatement matters because optimally located sensors is a central claim of the paper; please either prove the global property for this particular objective or soften the remark to describe the bound in Eq. (14) as an observability criterion that is tested numerically.","section":"Section 4.3, Remark"}],"minor_comments":[{"comment":"The keyword list contains the typo 'Inverse Problemas' and should read 'Inverse Problems'.","section":"Keywords"},{"comment":"The figure states 'Sensor size (Voxel) = 0.2 x 0.2 cm', while the text of Section 5.1 says pixels of size 2 x 2 cm^2; the units and numbers should be made consistent.","section":"Figure 8"},{"comment":"The reported mesh parameters appear inconsistent: hmin,2 = 7.2e-3 is larger than hmax,2 = 1.8e-3, which would make the stated minimum larger than the stated maximum for the second mesh.","section":"Section 3.3"},{"comment":"The word 'enumerator' should be 'numerator' in the sentence discussing the bound.","section":"Section 4.3, Eq. (14)"},{"comment":"The text alternates between m = 52 and m1 = 56 for the smallest sensor set; the notation should be unified, and the caption 'Greedy A.(m2 = 162)' appears incomplete.","section":"Sections 5.4 and Figures 15-16"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely publishable after the claims are narrowed and the bound is corrected. The main risk is overclaiming physical accuracy from self-consistency tests; the authors already acknowledge 'proof-of-concept' in the text, so a revision that aligns the abstract and conclusions with that framing is feasible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does what it says: it builds a PBDW-style reduced-order state estimator for a freezing salmon slice, adds a greedy sensor placement criterion based on the a-priori error bound, and tests the pipeline across several reconstruction scenarios. The ingredients are known, but the combination for a phase-change conjugate heat-transfer problem is new, and the implementation looks careful. The forward solver is standard but solid, and the validation against two benchmarks (beef freezing, air cavity) gives the numerics some external anchor.\n\nThe reconstruction studies are thorough: full-domain thermocamera-style measurements, food-blind extrapolation from the airflow region, decreasing sensor counts, and a test with real thermocouple data. The reported ~1% field errors are internally consistent and the greedy placement does beat regular grids. The code and data references are a plus.\n\nThe soft spots are real but proportionate. First, the main validation is synthetic: ground truth comes from the same FVM solver that generated the training snapshots, so the errors measure self-consistency, not physical accuracy. The authors are upfront about this—they call it a proof-of-concept and list LES as future work to strengthen credibility—so it is a scope limitation, not a hidden flaw. Second, the forward model is validated at Ra ~1e7–1e9 while the operating regime is declared as Ra > 1e10. If the URANS closure is biased there, the reconstruction inherits that bias because it projects onto states from the same model. The experimental test in Section 5.2 is at low Ra and uses a single-parameter ROM, so it does not close that gap. Third, the greedy optimality remark is overstated: the claim that β(m*) ≥ β(msub) for any other selection is not what a greedy algorithm guarantees. The numerical comparison is credible, but the language should be softened to a heuristic with strong empirical support. Minor: Eq. (13) mixes V and \\hat{V} in the SVD derivation; the result survives, but the notation needs cleaning.\n\nWho gets value from this: people working on ROM-based data assimilation, sensor placement, and food-engineering process monitoring. It is not a breakthrough, but it is a solid proof-of-concept with reproducible software and honest limitations. I would send it to peer review, with requests to add a high-Rayleigh validation or sensitivity discussion, test the external-only extrapolation against experimental or higher-fidelity data, and moderate the greedy optimality claims.","headline":"Competent proof-of-concept for ROM-based state estimation in food freezing, worth refereeing despite the synthetic-validation ceiling on its practical claims.","tokens_in":28957,"tokens_out":2741,"would_cite":true,"duration_ms":25903,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M32","65M08","80A22","76F99"],"pacs":[],"model":"deepseek-v4-flash","headline":"A physics-informed reduced-order inverse framework, trained on turbulent freezing simulations, reconstructs the internal temperature field of a freezing salmon slice from sparse external sensors with errors near 1 percent.","keywords":["food freezing","temperature estimation","data assimilation","reduced order modeling","optimal sensor placement","greedy algorithm","phase change","turbulent convection"],"falsifier":"Place thermocouples at several depths inside a real salmon slice, run the freezer with the same boundary-condition range as the paper, mount the greedy-chosen sensors in the airflow, and compare the reconstructed internal temperature evolution against the thermocouple record; the framework is validated only if the reconstructed temperatures track the thermocouples to within the same few percent it achieves against its own simulation. A complementary check in simulation is to scramble the sensor positions from the greedy layout: if random layouts yield errors as low as the greedy one, the observability maximization is not carrying the accuracy claim.","tokens_in":28014,"feed_emoji":"❄️","tokens_out":17299,"duration_ms":130994,"temperature":0.7,"pith_summary":"This paper tries to establish that the temperature field inside food during freezing can be reconstructed in real time from a small set of external temperature sensors, without ever touching the product. The authors combine a turbulent airflow simulation of a salmon slice in a freezer cabinet with a reduced-order model built from snapshots of that simulation, then recover the full temperature field by projecting sparse, low-resolution measurements onto the low-dimensional basis. A greedy algorithm picks sensor locations in advance, maximizing how observable the reduced-order dynamics are from those locations. In synthetic tests, the recovered fields match the simulated ground truth with relative errors around 1% — even when the sensors are placed only in the airflow and the food itself is unobserved. The value of the claim, if it holds with real measurements, is continuous quality-relevant monitoring of freezing processes at negligible online computational cost.","feed_headline":"Sparse airflow sensors read a freezing salmon's interior to ~1%","feed_subtitle":"Sparse airflow measurements become real-time freezing temperature maps via a greedy-chosen sensor layout.","key_machinery":"The central object is the reduced-order basis $\\Phi$ — the first $n$ left singular vectors of a snapshot matrix built from 48 full-order turbulent freezing simulations — together with the cross-Gramian matrix $G = W^{\\mathsf{T}}\\Phi$ that couples it to the sensors. Each sensor is a column of the observation matrix $W$, a normalized indicator function (Riesz representer) over a measurement pixel, so measurements are local averages of the temperature field. The reconstruction solves the normal equations $G^{\\mathsf{T}}G c = G^{\\mathsf{T}}\\ell$ in the $n$-dimensional space spanned by $\\Phi$; well-posedness requires the number of measurements $m$ to exceed the ROM dimension $n$, and the error is controlled by the smallest singular value of $G$. The greedy algorithm places the next sensor to maximize that smallest singular value, equivalently to shrink the a priori bound $e(n) = \\hat{S}_n^{-1} \\, (\\sum_{i>n} \\sigma_i^2 / \\sum_i \\sigma_i^2)^{1/2}$, so the sensor layout is computed once, offline, without any dependence on the measurement values.","core_discovery":"On the paper's own terms, the central claim is that a physics-informed reduced-order model — the first ~111 singular vectors of a 48-simulation snapshot set — together with a greedy sensor layout that maximizes the observability of that model, turns a severely underdetermined inverse problem into a well-posed one whose solution is the full temperature field. The quantified results are that the optimal 111-mode ROM reconstructs the temperature field over 16 held-out freezing simulations with peak errors near 5% at the start of freezing and errors below 1% for most of the process; that a greedy layout of 162 airflow sensors matches the accuracy of a regular 352-sensor array; and that restricting all sensors to the airflow, away from the food, still yields internal temperature errors of roughly 4 to 5%. The same reconstructed field supports derived quantities such as local freezing curves and freezing rates. The authors present the method as a proof of concept on synthetic data, with the full workflow — forward simulation, ROM training, sensor placement, and online reconstruction — designed to be transferable to experimental measurements.","pith_inferences":["Beyond the paper: the same observability criterion — maximize the smallest singular value of the sensor-to-ROM cross-Gramian — should transfer to other conjugate heat-transfer monitoring tasks (thawing, baking, pasteurization, cryopreservation) where a snapshot dataset and a candidate sensor pool exist, so a natural extension is to test the pipeline on a different food geometry or a different phas","Beyond the paper: because the reported ~1% error is measured against the URANS ground truth, the practical accuracy of the method in a real freezer hinges on the forward model's fidelity; an immediate testable extension is to repeat the reconstruction with experimental thermocouple data inside the salmon and report the error against those sensors instead of against the simulation.","Beyond the paper: the error bound (14) suggests an adaptive experimental-design loop — add sensors only where the smallest singular value of $G$ would grow most — which the paper does not explore; this could yield layouts that respond to changing flow regimes or food loads."],"forward_implications":["Real-time monitoring becomes feasible: once the offline basis and sensor layout are built, each temperature reconstruction is just the solution of an $n \\times n$ linear system with $n \\approx 111$, so new measurements can be assimilated essentially as they arrive.","Non-invasive quality control: temperature inside the food can be estimated from sensors in the airflow only, which is what makes the method practical for industrial freezers where probing the product is undesirable.","Sensor economy: a greedy layout of 162 sensors matches the reconstruction accuracy of a regular 352-sensor thermo-camera array, and sparser layouts with roughly 56 sensors still keep time-averaged errors comparable to full-domain measurements.","Derived quantities follow from the reconstructed field: local freezing curves and freezing rates can be computed at any point in the food, with local temperature errors below about 2% at control points, so quality-relevant indicators become available without extra sensors.","The sensor placement is reusable: because the greedy optimization depends only on the ROM and the candidate sensor pool, not on the measurement values, the same layout serves any future freezing run in the same geometry without repeating the optimization."],"supporting_citations":[{"why":"Supplies the parameterized-background data-weak (PBDW) formulation that the paper's regularized least-squares reconstruction instantiates.","marker":"[52]"},{"why":"Provides the perfect-model (no-bias) explanation of the ROM-based inverse estimator and the a priori bound the greedy criterion builds on.","marker":"[35]"},{"why":"Demonstrates the same ROM-based state-estimation and sensor-placement pipeline in 3D problems, the methodological template applied here to freezing.","marker":"[54]"},{"why":"Provides the implementation details of the observation matrix (Riesz representers) and the greedy/PBDW machinery this paper adapts.","marker":"[56]"},{"why":"Gives the greedy algorithm theory for optimal measurement selection in reduced-model state estimation that the paper adapts to sensor placement.","marker":"[62]"},{"why":"Validates the 2D turbulence-model-based freezing simulation of salmon in a freezer cabinet, justifying the forward model used to generate ground truth data.","marker":"[65]"},{"why":"Previous CFD solver for conjugate freezing with experimental data; used as the comparison benchmark for the P1 direct-problem validation.","marker":"[10]"},{"why":"Experimental benchmark for turbulent natural convection in a square cavity used to validate the URANS solver on P2.","marker":"[87]"},{"why":"Defines the k-omega SST turbulence model that closes the forward URANS equations.","marker":"[69]"}],"fun_headline_variants":["Greedy sensor layout sees into freezing food with ~1% error","ROM plus greedy sensors maps freezing salmon's interior in real time","162 airflow sensors match a 352-sensor array for freezing food monitoring","Physics-informed reduced-order model turns sparse sensors into full freezing maps","Optimal sensor placement reveals internal freezing temperatures in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the forward URANS simulation — its k-omega SST turbulence closure, effective-heat-capacity phase change, and 2D geometry — faithfully represents the real freezing process at the operating Rayleigh numbers (Ra $> 10^{10}$); every reported reconstruction error is measured against that simulation's output, so any physical mismatch becomes reconstruction error.","fun_headline_variants_meta":{"raw":{"variants":["Greedy sensor layout sees into freezing food with ~1% error","ROM plus greedy sensors maps freezing salmon's interior in real time","162 airflow sensors match a 352-sensor array for freezing food monitoring","Physics-informed reduced-order model turns sparse sensors into full freezing maps","Optimal sensor placement reveals internal freezing temperatures in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001691,"raw_usage":{"total_tokens":6709,"prompt_tokens":964,"completion_tokens":5745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":5656}},"tokens_in":580,"tokens_out":5745,"duration_ms":34169,"temperature":1.0,"reasoning_tokens":5656,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:38:14.058842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place thermocouples at several depths inside a real salmon slice, run the freezer with the same boundary-condition range as the paper, mount the greedy-chosen sensors in the airflow, and compare the reconstructed internal temperature evolution against the thermocouple record; the framework is validated only if the reconstructed temperatures track the thermocouples to within the same few percent it achieves against its own simulation. A complementary check in simulation is to scramble the sensor positions from the greedy layout: if random layouts yield errors as low as the greedy one, the observability maximization is not carrying the accuracy claim.","supporting_citations":[{"cited_title":"A parameterized-background data-weak approach to variational data assimilation: formulation, analysis, and application to acoustics,","cited_arxiv_id":null,"evidence_quote":"Supplies the parameterized-background data-weak (PBDW) formulation that the paper's regularized least-squares reconstruction instantiates."},{"cited_title":"A real-time variational data assimilation method with data-driven model enrichment for time-dependent problems,","cited_arxiv_id":null,"evidence_quote":"Provides the perfect-model (no-bias) explanation of the ROM-based inverse estimator and the a priori bound the greedy criterion builds on."},{"cited_title":"Fast reconstruction of 3D blood flows from Doppler ultrasound images and reduced models,","cited_arxiv_id":null,"evidence_quote":"Demonstrates the same ROM-based state-estimation and sensor-placement pipeline in 3D problems, the methodological template applied here to freezing."},{"cited_title":"Displacement and pressure reconstruction from magnetic resonance elastography images: Application to an In Silico brain model,","cited_arxiv_id":null,"evidence_quote":"Provides the implementation details of the observation matrix (Riesz representers) and the greedy/PBDW machinery this paper adapts."},{"cited_title":"Greedy algorithms for optimal measurements selection in state estimation using reduced models,","cited_arxiv_id":null,"evidence_quote":"Gives the greedy algorithm theory for optimal measurement selection in reduced-model state estimation that the paper adapts to sensor placement."},{"cited_title":"Energy analysis of convective freezer cabinet with PCMs and salmon-fillet during charging, discharging and normal operation processes by CFD modeling,","cited_arxiv_id":null,"evidence_quote":"Validates the 2D turbulence-model-based freezing simulation of salmon in a freezer cabinet, justifying the forward model used to generate ground truth data."},{"cited_title":"Experimental benchmark data for turbulent natural convection in an air filled square cavity,","cited_arxiv_id":null,"evidence_quote":"Experimental benchmark for turbulent natural convection in a square cavity used to validate the URANS solver on P2."},{"cited_title":"Ten years of industrial experience with the sst turbulence model,","cited_arxiv_id":null,"evidence_quote":"Defines the k-omega SST turbulence model that closes the forward URANS equations."}],"review_version":1}