{"id":"3bac6efa-a59c-4559-8352-158a77a4de7d","arxiv_id":"2507.09356","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A PINN recovers jet-impingement convective heat transfer coefficients from sparse, noisy in-solid temperatures, matching CHT-based benchmarks with relative errors below 8% for noise up to 10% and sampling rates at or above 0.5 s^{-1}.","lead":"This paper tests a physics-informed neural network that estimates how strongly a cooling air jet removes heat from a metal plate using only sparse temperature readings inside the plate. It reports errors below 8 percent when noise and sampling rates stay within a tested range, checked against a detailed conjugate heat transfer simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quadratic ansatz for h_fc(x) predetermines the recovered profile; the sub-8% errors largely reflect this parametric bias and do not support the paper's general claim of estimating arbitrarily spatially varying CHTCs.","rationale":"The reader's weakest-assumption analysis identifies the second-degree polynomial restriction on h_fc(x), and I agree that this is the load-bearing issue. The parametric form is the main reason the inverse problem is tractable with only eight sensors: a three-parameter fit can be well constrained, but it cannot represent sharper stagnation peaks or secondary maxima. The paper's reported sub-8% errors are relative to a profile generated by the same CHT simulation, so they are internally consistent, but the zero-noise error floor of about 6% indicates that the method is essentially fitting the assumed polynomial shape. The abstraction and conclusions generalize this to real-world jet impingement cooling without evidence that the functional class is adequate. This is an external-validity and overclaiming concern rather than an internal logical inconsistency, so it does not invalidate the numerical demonstration; it should be addressed by reframing the claims or by testing on more general profiles. The reader's CONDITIONAL verdict already captures this, so I do not recommend a verdict change.","tokens_in":19795,"tokens_out":5653,"duration_ms":73948,"concrete_test":"Compute the L2 projection of the CHT benchmark h_fc(x) profile onto the space of quadratic polynomials. If the projection error is close to the zero-noise entries in Table 4 (~6%), then the PINN is recovering only the anticipated quadratic component. Independently, generate synthetic data from a CHT profile with a secondary maximum (by varying H/D or using a literature jet-impingement profile) and retrain the same architecture; if the relative error at 10% noise and 0.5 s^-1 sampling exceeds 8%, the generality claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is in Section 5.2: h_fc(x) is restricted to a second-degree polynomial. This reduces the spatially varying inverse problem to three parameters, and the reported errors are measured against a profile that the ansatz can approximate, not against the full CHT profile. Internal evidence: Table 4 shows zero-noise relative errors of about 6.1-6.4% at every sampling rate. This noise-free error floor is the L2 distance between the CHT benchmark and its best quadratic approximation, so even with perfect data the method cannot improve. All sub-8% noisy-data claims inherit this bias. Real impinging-jet CHTC profiles can have sharper stagnation peaks and secondary maxima (e.g., at r/D about 1.5-2), which a quadratic cannot represent. The paper's broad statement (Abstract and Section 6) that the framework is a scalable alternative for real-world cooling applications therefore rests on an untested smoothness assumption. The comparison is internally consistent because the benchmark and the synthetic data both come from the same CHT simulation, but it does not test recovery of profiles outside the assumed functional class.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a physics-informed neural network (PINN) framework for the inverse estimation of the convective heat transfer coefficient (CHTC) at the fluid–solid interface in a jet impingement configuration at Reynolds number 5000. Temperature data are generated synthetically from a conjugate heat transfer (CHT) simulation, and the PINN solves the 2D transient heat conduction equation in the solid while treating the unknown boundary coefficient as a trainable parameter. Two cases are considered: a constant, spatially averaged CHTC and a spatially varying CHTC parameterized as a quadratic polynomial. The method is tested under additive Gaussian noise levels from 0% to 30% and sampling rates from 0.25 to 4.0 s−1, with relative L2 errors computed against CHT-derived benchmarks. The central quantitative claim is that for noise up to 10% and sampling rates of 0.5 s−1 or higher, the estimated CHTCs have relative errors below 8%.","tokens_in":20066,"tokens_out":6948,"duration_ms":73335,"significance":"If the central claim is valid, the paper offers a useful demonstration that a PINN can recover a boundary heat-transfer coefficient from sparse, noisy in-solid temperature measurements without modeling the fluid domain. The study has concrete strengths: it uses a non-dimensional formulation, systematically tabulates errors over a noise/sampling grid, includes a grid-independence study for the CHT benchmark, and reports implementation details (DeepXDE, optimizer, architecture). The conditional claim is consistent with Tables 3 and 4 for the tested cases. However, because the spatially varying CHTC is restricted to a quadratic polynomial, the validation is in-distribution: it does not test recovery of profiles outside this functional class, and the zero-noise error floor in Table 4 (6.1–6.4%) shows that the method cannot improve beyond the best quadratic approximation of the benchmark. The significance is therefore real but narrower than the abstract suggests.","major_comments":[{"comment":"The spatially varying inverse problem is reduced to fitting three polynomial coefficients because h_fc(x) is assumed to follow a second-degree polynomial. At zero noise, Table 4 reports relative errors of 6.1–6.4% at every sampling rate; this floor is the L2 approximation error between the CHT benchmark and its best quadratic fit, so it is not reduced by more data or higher sampling. The subsequent sub-8% results for noise up to 10% therefore demonstrate recovery of a quadratic-in-class profile, not estimation of a general spatially varying CHTC. The broad statements in the Abstract and Section 6 that the framework is a scalable alternative for real-world cooling applications overstate the evidence unless the authors either restrict the claim to profiles well represented by a quadratic or validate against benchmark profiles with sharper stagnation peaks or secondary maxima outside this class.","section":"Section 5.2, Table 4"},{"comment":"The noise model is specified only as 'additive Gaussian noise ranging from (0-30)%'; the manuscript does not state whether the noise standard deviation is a percentage of the local temperature, of the full temperature range, of the sensor full scale, or of some other reference value. Without this definition, the noise-level thresholds (e.g., 'up to 10%') used in Tables 3–4 and the Abstract cannot be reproduced or compared across configurations. Please specify the noise generation formula, report the actual signal-to-noise ratio, and state whether independent noise is added to each sensor time series.","section":"Section 4.1, 'Data acquisition from synthetic measurement'"}],"minor_comments":[{"comment":"The Abstract says 'relative errors below 7.6%' for noise up to 10% and sampling rates of 0.5 s−1 or higher, but Table 4 at 10% noise and 0.5 s−1 reports 7.624%; this value is not strictly below 7.6%. The wording should be reconciled with the 'within 8%' statement in Section 6.","section":"Abstract; Table 4"},{"comment":"The sentence 'As shown in Table 3' appears when discussing the spatially varying CHTC case; the referenced table should be Table 4 rather than Table 3.","section":"Section 5.2"},{"comment":"The caption of Figure 1 describes a PINN architecture, but the figure appears to show the jet impingement geometry; the caption should be corrected, since the PINN architecture is displayed in Figure 3.","section":"Figure 1"},{"comment":"The Nomenclature defines α as thermal diffusivity, but Table 2 lists 'Thermal Conductivity αs: 16.27 W m−1 K−1'; the symbol for thermal conductivity should be κ_s, and the thermal diffusivity should be reported separately with consistent units.","section":"Table 2 and Nomenclature"},{"comment":"There are typographical errors that should be corrected, including 'axis-symmetric' instead of 'axisymmetric' in Section 2.2 and 'follwoing' in the definition of the initial loss in Section 3.","section":"Section 2.2 and Section 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a fluids/thermal-fluids journal, and the conditional numerical claim is broadly consistent with the reported tables. The main weakness is the quadratic functional ansatz for the spatially varying CHTC, which makes the validation in-distribution; the authors should either reframe the claims to the quadratic class or add a test case with a profile outside that class (e.g., with a sharper stagnation peak or a secondary maximum). The noise definition should also be made precise. I do not see a circularity problem: the CHT benchmark is independent of the PINN's fitted parameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper's central claim is narrower than the abstract implies. The spatially varying CHTC is assumed to be a second-degree polynomial, so the sub-8% errors measure how well the PINN recovers three polynomial coefficients, not how well it recovers arbitrary heat transfer profiles. With that caveat, the study is a clean, well-documented synthetic demonstration. What is genuinely useful: the systematic noise and sampling-rate study for jet impingement CHTC inversion using PINNs, benchmarked against CHT simulations. The constant-coefficient case is straightforward but solid, with errors mostly below 5% even at moderate noise. The spatially varying case shows the expected degradation with noise and the benefit of higher sampling rates. The tables support the internal claims, and the grid independence study in the appendix is a good check. The paper also gives enough detail on architecture and training that someone could reproduce the setup. The soft spots are real but not disqualifying. The quadratic ansatz is the big one. Table 4 shows a noise-free error floor around 6.1-6.4% at every sampling rate, which is essentially the distance between the CHT benchmark and its best quadratic fit. So the reported accuracy is inherited from the smoothness assumption, not from the PINN's inferential power. Real impinging-jet CHTC profiles can have secondary maxima and sharper stagnation peaks that a quadratic cannot represent, and the paper does not test any profile outside that function class. The conclusion that the framework is a scalable alternative for real-world applications overreaches. Second, the paper claims an advantage over traditional inverse heat transfer methods but never runs one. No Tikhonov or function-specification baseline appears anywhere. That makes the comparative claim unsubstantiated, even if the absolute errors look fine. Third, the noise definition is vague. Additive Gaussian noise ranging from 0-30% is not tied to a signal-to-noise ratio or temperature scale, so the noise levels are not precisely interpretable. Also, no code or data is provided, so the numbers cannot be independently checked. Overall, this is a competent engineering case study, not a methodological breakthrough. It deserves a serious referee who can push for a sharper framing and a classical baseline. I would not cite it in my own work, but I would send it to review and probably accept after moderate revision if the authors tone down the general claims and add a traditional IHTP comparison.","headline":"Credible PINN-based CHTC inversion demo, but the sub-8% claim is conditional on a quadratic profile ansatz, and there is no comparison to classical inverse methods.","tokens_in":729,"tokens_out":871,"would_cite":false,"duration_ms":27292,"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":"A physics-informed neural network recovers jet-cooling heat transfer coefficients from sparse in-solid temperatures, matching conjugate heat transfer benchmarks within 8% at noise up to 10% and sampling at 0.5 per second or faster.","keywords":["jet impingement cooling","physics-informed neural networks","inverse heat transfer","conjugate heat transfer","convective heat transfer coefficient","Biot number","transient heat conduction","sparse noisy temperature data"],"falsifier":"Generate synthetic conjugate heat transfer data from a deliberately non-polynomial coefficient profile — for example, a Gaussian stagnation peak narrower than any second-degree polynomial, or a profile with an off-center secondary maximum — and run the paper's identical PINN protocol on it. If the recovered profile inherits the polynomial bias and the relative $L^2$ error exceeds 8%, then the claim is shown to hold for polynomial-shaped profiles only, which bounds the method's scope to configurations whose true coefficient profile a second-degree polynomial can represent.","tokens_in":19609,"feed_emoji":"🌡️","tokens_out":28423,"duration_ms":253557,"temperature":0.7,"pith_summary":"A physics-informed neural network can estimate the convective heat transfer coefficient (CHTC) at the cooled surface of a jet-impinged metal plate from sparse, noisy temperature readings taken inside the solid, with no simulation of the fluid flow at all. If the result holds, engineers could extract a key cooling-design parameter from a few embedded thermocouples instead of running expensive conjugate heat transfer (CHT) simulations or dense instrumented experiments. The paper validates the method on synthetic data from high-fidelity CHT simulations of a round air jet at Reynolds number 5000, treating the simulation as ground truth. With additive Gaussian noise up to 10% and sampling at 0.5 readings per second or faster, the estimated CHTC—both a single averaged value and a spatially varying profile—matches the benchmark at relative errors below 8%. The mechanism is to treat the unknown boundary coefficient as a trainable parameter while enforcing the transient heat conduction equation as a soft constraint in the network's loss.","feed_headline":"Recover jet-cooling heat-transfer coefficients within 8 percent","feed_subtitle":"The network uses only sparse in-solid temperature readings, so no fluid simulation is needed.","key_machinery":"The carrying object is the composite loss $L = L_{\\mathrm{data}} + L_{\\mathrm{PDE}} + L_{\\mathrm{boundary}} + L_{\\mathrm{initial}}$, in which the unknown heat transfer coefficient enters only through the upper-boundary Robin condition $\\partial T/\\partial y + \\mathrm{Bi}_{fc}\\,T = 0$. The PINN is a fully connected feedforward network taking $(x,y,t)$ as input; the unknown appears as additional trainable variables — a single $\\mathrm{Bi}_{fc}$ for the averaged case, or the polynomial coefficients of $\\mathrm{Bi}_{fc}(x)$ for the spatially varying case — so one network solves the forward heat equation and the inverse parameter-estimation problem at once. Temperature residuals at the eight thermocouple sites drive the data term, while the PDE, boundary, and initial terms keep the field physical; nondimensionalizing by plate radius and initial temperature difference converts the unknowns into Biot numbers and keeps all data in the unit range, which the paper credits with stabilizing training.","core_discovery":"The central claim, stated on the paper's own terms, is that the convective heat transfer coefficient at the fluid–solid interface is identifiable from interior temperature measurements through a physics-informed neural network that embeds the nondimensionalized two-dimensional transient heat conduction equation, its initial condition, and its convective boundary conditions in a composite loss. In the simpler test case the unknown is a single constant Biot number representing the spatially averaged coefficient; in the harder case it is the three coefficients of a second-degree polynomial $h_{fc}(x)$, which the paper argues suits axisymmetric jet impingement profiles. Using eight synthetic thermocouples placed just below the impingement surface, the network recovers the stagnation-region peak and its radial decay, and the reported relative $L^2$ errors stay below 8% for noise up to 10% and sampling rates of 0.5 s$^{-1}$ or higher. At noise levels up to 30%, accuracy degrades but is partially restored by denser sampling, and the paper reports that moderate noise can occasionally improve generalization through an implicit regularization effect.","pith_inferences":["The 8% accuracy claim is, strictly, a claim about how well a second-degree polynomial can represent the benchmark profile; real jet profiles with sharper stagnation peaks or secondary maxima would exercise the framework beyond its stated parametric scope.","The method presumes the solid's thermal properties and the side-wall free-convection coefficient are known exactly; the effect of getting those inputs wrong is not tested, so field application to an unknown alloy or non-axisymmetric geometry would need a separate sensitivity study.","The same construction — a trainable boundary coefficient inside a physics-constrained loss — plausibly transfers to other boundary parameters such as heat flux, contact resistance, or surface heat sources, and the paper's noise-versus-sampling-rate trade-off provides a template for choosing sensor cadence before deployment.","The eight thermocouples are fixed just below the surface, so the reported errors are tied to that sensor layout; the paper lists optimal sensor placement as future work, implying the 8% figure will shift with probe depth and spacing."],"forward_implications":["One trained network replaces a full CHT simulation: each inverse run took about 50,000 iterations and roughly 6 minutes on the reported hardware, which is the practical cost of obtaining the boundary coefficient.","The same framework covers both a lumped averaged coefficient and a spatially varying profile, so a user can choose modeling fidelity without changing the network or the data acquisition.","When noise is high, sampling rate is the lever: at 30% noise the spatially varying profile error drops from about 50% at 0.25 s$^{-1}$ to under 10% at 4.0 s$^{-1}$, so time resolution compensates for sensor imprecision.","The method needs no fluid-domain knowledge, which is exactly the situation in which CHT simulation is impractical and surface measurement is intrusive.","For the varying-coefficient case the recovered profile reproduces the expected stagnation-peak shape, confirming that the approach captures spatially local heat-transfer behavior rather than only a global average."],"supporting_citations":[{"why":"Introduces the physics-informed neural network method whose composite-loss construction this paper adopts for inverse estimation.","marker":"[29]"},{"why":"Supplies the experimental jet impingement geometry and operating conditions that the CHT simulations replicate, anchoring the synthetic data to a measured configuration.","marker":"[34]"},{"why":"Provides the software library used to implement and train the PINN models.","marker":"[35]"},{"why":"Defines the classical inverse heat transfer problem class that the paper's inverse formulation extends and contrasts with.","marker":"[16]"},{"why":"Extends PINNs to direct and inverse heat conduction problems of materials, the immediate methodological precedent for this paper's setup.","marker":"[33]"},{"why":"Supports the nondimensionalized formulation that the paper credits with improving PINN training stability.","marker":"[32]"},{"why":"Exemplifies the conjugate heat transfer modeling of impingement jets that serves as the benchmark approach for the estimated coefficients.","marker":"[14]"},{"why":"Is invoked to explain why moderate noise sometimes improves the estimated coefficients through implicit regularization.","marker":"[39]"}],"fun_headline_variants":["No fluid sim needed: PINN learns jet cooling from solid temps","PINN recovers jet impingement heat transfer from sparse solid data","Jet cooling coefficients from 8 sensors with <8% error","Physics-informed NN estimates jet cooling heat transfer without CFD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the spatially varying heat transfer coefficient is exactly a second-degree polynomial in position along the surface; a real jet profile with a sharper stagnation peak or secondary humps could not be expressed by that assumed shape, and no amount of data or sampling would remove the resulting bias.","fun_headline_variants_meta":{"raw":{"variants":["No fluid sim needed: PINN learns jet cooling from solid temps","PINN recovers jet impingement heat transfer from sparse solid data","Jet cooling coefficients from 8 sensors with <8% error","Physics-informed NN estimates jet cooling heat transfer without CFD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000666,"raw_usage":{"total_tokens":3099,"prompt_tokens":1064,"completion_tokens":2035,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":1963}},"tokens_in":680,"tokens_out":2035,"duration_ms":16021,"temperature":1.0,"reasoning_tokens":1963,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:58:26.227962+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic conjugate heat transfer data from a deliberately non-polynomial coefficient profile — for example, a Gaussian stagnation peak narrower than any second-degree polynomial, or a profile with an off-center secondary maximum — and run the paper's identical PINN protocol on it. If the recovered profile inherits the polynomial bias and the relative $L^2$ error exceeds 8%, then the claim is shown to hold for polynomial-shaped profiles only, which bounds the method's scope to configurations whose true coefficient profile a second-degree polynomial can represent.","supporting_citations":[{"cited_title":"Woodbury, and Gyula Gr´ of","cited_arxiv_id":null,"evidence_quote":"Introduces the physics-informed neural network method whose composite-loss construction this paper adopts for inverse estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the experimental jet impingement geometry and operating conditions that the CHT simulations replicate, anchoring the synthetic data to a measured configuration."},{"cited_title":"Deep- XDE: A Deep Learning Library for Solving Differential Equations","cited_arxiv_id":null,"evidence_quote":"Provides the software library used to implement and train the PINN models."},{"cited_title":"Numerical study of transient con- jugate heat transfer of a turbulent impinging jet","cited_arxiv_id":null,"evidence_quote":"Defines the classical inverse heat transfer problem class that the paper's inverse formulation extends and contrasts with."},{"cited_title":"Witschey, John A","cited_arxiv_id":null,"evidence_quote":"Supports the nondimensionalized formulation that the paper credits with improving PINN training stability."},{"cited_title":"Bountin, Boris V","cited_arxiv_id":null,"evidence_quote":"Exemplifies the conjugate heat transfer modeling of impingement jets that serves as the benchmark approach for the estimated coefficients."},{"cited_title":"Fern´ andez-Mart ´ ınez, J.L.G","cited_arxiv_id":null,"evidence_quote":"Is invoked to explain why moderate noise sometimes improves the estimated coefficients through implicit regularization."}],"review_version":1}