{"id":"f371b6f4-1ac2-4d7c-b0ec-d7e288150489","arxiv_id":"2509.05886","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A self-supervised physics-informed neural network is claimed to estimate liquid sodium Nusselt numbers in miniature heat sinks to about 8% error, but the evidence for an advantage over plain neural networks is weak and the network's loss function is ambiguously defined.","lead":"A machine learning study tests several regression methods, including a physics-informed neural network, to predict heat transfer of liquid sodium in miniature heat sinks from 87 CFD simulations. The claimed best method, SPINN, estimates Nusselt numbers within about 8% error, but the paper's own numbers show a plain neural network had lower cross-validation error.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) as written cannot implement the claimed physics coupling: the MSE term compares a sigmoid-bounded physics coefficient [0,1] with a Nusselt prediction of ~2.2–3.85, so the loss is unsatisfiable and the 'physics-informed optimal' claim is unsupported.","rationale":"The reader's weakest assumption and my independent reading converge on Eq. (12): the physics term is never specified as a function of the network output or as a bounded quantity matched to the output scale. Since the loss is the only mechanism connecting physics to learning, an unsatisfiable MSE term makes the 'physics-informed' enhancement vacuous. I also checked the paper's internal consistency: Table 5 lists PINN MAPE 0.0185 against NN 0.0028 and transfer 0.0020, so the word 'optimal' in the title cannot be defended on the paper's own primary error metric. The holdout-based 'strictly within 8%' claim is further weakened by the same section saying all methods are mostly within +8%, and by the holdout being a small subset of 87 points. These are internal inconsistencies, not disagreements with outside consensus, so they are directly disqualifying for the central claim. I do not see a need to move the reader's verdict: the manuscript remains interesting as a dataset-specific comparison of SVR/GP/NN, but not as a valid demonstration of an optimal self-supervised PINN framework. The proposed ablation is the minimal check that would determine whether the physics term has any causal effect, and the scale-mismatch check is a direct verification of the mathematical objection.","tokens_in":12323,"tokens_out":7507,"duration_ms":70953,"concrete_test":"Reimplement the architecture and Eq. (12) exactly as described (3 hidden layers with 20/20/12 neurons, sigmoid physics-coefficient head, Adam lr=0.34) on the 87 Na points, and log the physics MSE, the coefficient, and the Nu_hat_ave predictions each epoch. If the physics MSE remains large (>=2) and the coefficient saturates at ~1 while predictions stay ~2.2–3.85, Eq. (12) is unsatisfiable as written. Then run the same architecture with the MSE term removed: if the holdout max error and CV MAPE do not change materially, the physics term is not load-bearing; if removing it materially improves MAPE, it is actively harmful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3's central claim—self-supervised PINN is optimal and 'strictly within 8%'—rests on Eq. (12), whose physics term is ill-posed. Loss = MAPE(prediction, labels) + MSE(physics prediction, physics coefficient). The physics coefficient is a sigmoid neuron (Eq. 14), bounded to [0,1], while the 'physics prediction' is Nu_hat_ave from Eq. (11), which Table 1 shows lies between about 2.2 and 3.85. A bounded scalar cannot match that target, so the MSE term cannot be minimized to zero and the gradient it supplies does not push the Nusselt prediction toward Eq. (11). If Nu_hat_ave is a fixed formula, only the coefficient (and the hidden features feeding it) are pulled toward 1; if it is a network head, the loss pulls the output scale down toward [0,1]. Either way 'balancing data and physics' is not what Eq. (12) does. The paper's own Table 5 aggravates this: PINN's 10-fold CV MAPE (0.0185) is 6.6x worse than plain NN (0.0028), so the 'optimal' designation is internally contradicted; and the conclusion that 'only self-supervised PINN remains within the 8% range' contradicts the same section's statement that all methods are 'mostly within an 8% margin of error.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops machine-learning surrogate models for predicting the average Nusselt number of liquid sodium in miniature heat sinks from a small dataset of 87 CFD simulations. It compares Gaussian process, support vector regression, a shallow neural network, a transfer-learning variant, and a proposed 'self-supervised' physics-informed neural network (PINN) with an added physics-coefficient neuron. The central claims are that this SPINN framework is 'optimal', that it balances data and physics based on their uncertainty, that it achieves errors strictly within 8% on a holdout set, and that it is more robust than other ML methods.","tokens_in":12622,"tokens_out":4471,"duration_ms":38065,"significance":"If the method worked as described, the paper would offer a pragmatic surrogate for an expensive CFD problem and a possible template for combining empirical correlations with neural networks on small datasets. The paper deserves credit for a structured validation protocol: 10-fold cross-validation, Monte Carlo simulation with 500 evaluations, and systematic comparison across several algorithms. The application to liquid-sodium heat transfer is practically motivated. However, the central physics-informed mechanism is not correctly implemented, the 'physics' term is an empirical fit to the same dataset rather than an independent constraint, and the paper's own benchmark table contradicts the 'optimal' designation. The significance of the contribution as it stands is therefore low.","major_comments":[{"comment":"The loss function in Eq. (12) is ill-posed: it adds MSE(physics prediction, physics coefficient), where the physics prediction Nu_hat_ave from Eq. (11) lies between roughly 2.2 and 3.85 (Table 1), while the physics coefficient is a sigmoid neuron bounded to [0,1] by Eq. (14). This MSE term cannot be minimized to zero, and its gradient does not pull the Nusselt prediction toward the physics correlation. The claimed balancing of data and physics is therefore not implemented by the stated loss.","section":"Sec. 4.3, Eq. (12)"},{"comment":"The 'physics' used in the PINN is the empirical correlation of Pourghasemi and Fathi [12], which was fitted to the same or closely related CFD data that constitutes the training set. Using this correlation as a self-supervised constraint on a holdout from that same 87-point dataset is circular: the physics term is not an independent physical law. The abstract and conclusion claim the model 'adheres to physical laws', but Eqs. (8)-(11) are a curve fit, not governing equations such as (1)-(4).","section":"Sec. 2(e), Eqs. (8)-(11)"},{"comment":"The paper's central 'optimal' claim is internally contradicted by its own results: the 10-fold cross-validated MAPE of SPINN is 0.0185, about 6.6 times worse than the plain NN (0.0028) and worse than SVR-Bayesian (0.0125). Furthermore, the conclusion states that 'only self-supervised PINN remains within the 8% range', but the same section says all methods provide estimations 'mostly within an 8% margin of error', and Figure 11 shows all three NN-based methods inside the band. These statements are inconsistent.","section":"Table 5 and Sec. 4.3"},{"comment":"The manuscript says the physics-coefficient layer 'determines the weight of physics in the loss function to balance data and physics based on their uncertainty', but Eq. (12) is an unweighted sum of MAPE and MSE, and the coefficient is the target of the MSE, not a weight. The mechanism by which a scalar in [0,1] modulates the data-versus-physics trade-off, or is updated during backpropagation, is never specified. It is also unclear how the physics prediction Nu_hat_ave is computed for each input sample, since Eq. (8) appears garbled in the text.","section":"Sec. 2(e) and Fig. 3"}],"minor_comments":[{"comment":"There are many typographical errors and unclear phrases, including 'Besian' instead of 'Bayesian', 'neroun' for 'neuron', 'MAPPE' for 'MAPE', 'the weight the of physics', and a garbled Eq. (8). The paper needs careful proofreading before any further consideration.","section":"Throughout"},{"comment":"The text says the inputs are in R6, but only five inputs are listed (Pe, Dh, alpha, L, W). This inconsistency should be corrected or clarified.","section":"Sec. 2.1(c)"},{"comment":"The text states 'Figure 3 illustrates that transferring the first layer results in the lowest MAPE error', but the relevant figure appears to be Figure 6; the cross-reference is wrong.","section":"Sec. 4.2"},{"comment":"The size of the holdout dataset is never stated. The validation section refers to a holdout set, but the number of points and how it was carved from the 87-point dataset are not described, which hampers reproducibility.","section":"Sec. 2, Data preprocessing"},{"comment":"The statement that SPINN achieves an error 'strictly within 8%' is ambiguous because MAPE is reported as 0.0185 (1.85%), while Figures 10-11 use an 8% band around the parity line. The relation between these two error metrics should be made explicit.","section":"Sec. 4.3"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an incomplete working paper: the acknowledgments state that a revised version with co-authors is in preparation. The physics correlation used as 'physics' comes from the same group's prior work on the same dataset, which amplifies the circularity concern. The internal contradiction between the reported PINN error and the 'optimal' claim is sufficient for rejection on its own; the ill-posed loss function in Eq. (12) is a further load-bearing flaw that cannot be fixed by simple revision within the manuscript's current scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my take on this SPINN paper. The useful part is the 87-point CFD dataset for sodium Nusselt numbers in miniature heat sinks and the straightforward comparison of GP, SVR, NN, transfer learning, and a PINN variant on it. The transfer-learning finding—first-layer transfer from a water-trained NN gives MAPE 0.002 versus 0.0028 without—is modest but real. The Monte Carlo sensitivity checks on hyperparameters are also a reasonable validation practice. If the paper were just a benchmark of standard ML tools on a new dataset, it would be a decent niche contribution.\n\nThe headline claim does not hold up. Equation (12) defines the PINN loss as MAPE(prediction, labels) plus MSE(physics prediction, physics coefficient), where the physics coefficient is a sigmoid neuron bounded to [0,1] and the physics prediction is a Nusselt number on the order of 2.2–3.85. That MSE term is unsatisfiable as written and does not create the claimed coupling between the physics correlation and the output. The stress-test note is correct: the loss is ill-posed. Even if it were well-posed, the \"physics\" in Eq. (11) is an empirical correlation fitted in [12] by the same group to the same or closely related CFD data, so it is not an independent constraint. The paper's own Table 5 compounds the problem: the PINN's 10-fold CV MAPE is 0.0185, about 6.6 times worse than plain NN's 0.0028. Calling this framework \"optimal\" is unsupported.\n\nThere are also internal contradictions in the claims. The conclusion says \"only self-supervised PINN remains within the 8% range,\" but the results section earlier says all ML methods are \"mostly within an 8% margin.\" The writing is rough—typos like \"neroun\" and \"Besian\"—and several sentences read as a first draft.\n\nNet assessment: the dataset and transfer-learning comparison have some value, but the central PINN contribution is not a working method as presented. I would not send this out as-is. Desk reject with a clear message: drop the optimality claim, fix Eq. (12) or remove the PINN section, and resubmit as an empirical comparison. A serious referee could help only after those load-bearing issues are addressed.","headline":"Useful benchmark of ML surrogates for Na heat-sink Nusselt numbers, but the self-supervised PINN claim is unsupported by an ill-posed loss and by the paper's own numbers.","tokens_in":13246,"tokens_out":3725,"would_cite":false,"duration_ms":32906,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a self-supervised physics-informed neural network with a learned physics-coefficient neuron estimates liquid-sodium Nusselt numbers in miniature heat sinks strictly within an 8% error using only 87 CFD data points.","keywords":["physics-informed neural network","self-supervised learning","Nusselt number","liquid sodium","miniature heat sink","transfer learning","surrogate model","uncertainty quantification"],"falsifier":"A reader could settle the central claim by ablating the physics term: train the same architecture with only the MAPE loss and compare holdout error and variance; if removing the MSE physics term leaves the 8% bound and the low variance unchanged, the physics coefficient is not doing the work attributed to it. In addition, replacing Eq. (11) with an independent correlation and re-checking the 8% bound would test whether the physics term is an independent constraint.","tokens_in":12000,"feed_emoji":"🌡️","tokens_out":7087,"duration_ms":59174,"temperature":0.7,"pith_summary":"The paper aims to show that a small, physics-informed neural network can replace days-long CFD simulations for predicting how well liquid sodium transfers heat in miniature rectangular heat sinks. Using only 87 CFD-generated data points, it trains kernel methods, ordinary neural networks, a transfer-learned network, and a self-supervised physics-informed network, then compares their holdout errors. The central assertion is that the self-supervised PINN, whose loss couples a data-fidelity term with a physics-correlation term through a learned physics-coefficient neuron, stays strictly within an 8% error margin and is more robust to random weight initialization than ordinary neural networks. The paper also claims that transferring the first layer of a water-trained network to the sodium network slightly improves accuracy, and that all methods are mostly within 8% for this dataset. If true, the framework gives engineers a fast surrogate for designing liquid-metal-cooled heat sinks where experiments are scarce and CFD costs about three days per point.","feed_headline":"Physics-informed AI predicts sodium cooling within 8%","feed_subtitle":"Using just 87 CFD runs, the hybrid model beats physics-only regression and ordinary neural nets on holdout data.","key_machinery":"The load-bearing object is the self-supervised PINN architecture: a fully connected network with ReLU hidden units, a linear output for $Nu_{ave}$, a sigmoid physics-coefficient neuron, and a two-term loss. The first term is MAPE between predicted and CFD Nusselt numbers; the second is MSE between the coefficient and the value of an empirical heat-transfer correlation $\\hat{Nu}_{ave}$ (Eq. 11) evaluated from the inputs. The sigmoid keeps the coefficient in (0,1), and Bayesian optimization tunes the layer sizes and learning rate. The intended mechanism is that the coefficient learns how much to trust the physics correlation versus the data, so the network is self-supervised by the physics formula rather than by labeled physics outputs. Supporting machinery includes 10-fold cross-validation for hyperparameter selection, 500-run Monte Carlo simulation to measure variance of predictions and MAPE, and genetic-algorithm selection of which layers to transfer from a water-trained network.","core_discovery":"On the paper's own terms, the discovery is that a self-supervised PINN with a dedicated physics-coefficient neuron can estimate the average Nusselt number ($Nu_{ave}$) of liquid sodium in SS-316 rectangular miniature heat sinks within a strict 8% error on a holdout dataset, using a training set of just 87 CFD simulations. The network outputs both the Nusselt number and a scalar coefficient bounded between 0 and 1; the loss is MAPE between the output and the data plus MSE between the physics-based correlation $\\hat{Nu}_{ave}$ (Eq. 11) and that coefficient. The paper interprets this as balancing data and physics according to their uncertainty, and reports that the PINN has about half the prediction variance of ordinary NN (max variance 0.1146 vs 0.2323) while remaining within the 8% band. A second claimed result is that transfer learning from a water-trained shallow NN, specifically the first layer, reduces MAPE from 0.0028 to 0.0020 for sodium, confirming that the most general layer transfers best. The paper positions the combined data-physics approach as more robust than physics-only regression, which has 5% to 10% error, and calls the framework optimal for data-scarce heat transfer modeling.","pith_inferences":["A testable extension is to make the physics coefficient actually gate the output, for example $Nu_{pred} = c \\cdot \\hat{Nu}_{ave} + (1-c) \\cdot Nu_{NN}$; the current loss only compares the coefficient to the correlation, so the claimed physics-informed improvement may come from regularization rather than a hard physical constraint.","Because Eq. (11) was fitted to the same 87-point CFD dataset, the physics term is not an independent source of information; re-testing the method with a correlation derived from separate experiments or from first-principles dimensionless analysis would clarify whether the physics term adds real constraint.","The strict-within-8% claim rests on a single holdout from one CFD campaign; a natural stress test is to hold out entire flow regimes, such as all turbulent cases, and see whether the PINN extrapolates, or to apply the same architecture to another liquid metal such as NaK or lead-bismuth eutectic."],"forward_implications":["If the 8% holdout bound holds beyond this dataset, engineers can screen Na-cooled miniature heat sink designs with a neural surrogate instead of running multi-day CFD simulations.","The PINN's roughly halved prediction variance (0.1146 vs 0.2323 max variance) implies the physics term stabilizes the network against weight-initialization randomness, which matters when the model must be retrained for new geometries.","Transfer learning from water to sodium means a source-domain NN trained on an abundant fluid can jump-start a target-domain model for a scarce-data liquid metal, cutting training epochs and improving MAPE.","Since physics-only regression gives 5% to 10% error while the hybrid PINN goes under 8%, the results support ensemble-like combination of empirical correlations with ML rather than choosing one over the other."],"supporting_citations":[{"why":"Supplies the 87 CFD Nusselt-number data points and the empirical physics correlation (Eq. 11) that the PINN's physics loss term is built on.","marker":"[12]"},{"why":"Provides the CFD mesh and simulation framework for liquid-metal miniature heat sinks that generated the training data.","marker":"[17]"},{"why":"Defines the water-trained shallow NN structure and its MAPE, which transfer learning adapts to sodium.","marker":"[14]"},{"why":"Supports the claim that kernel-based and NN methods perform well on small datasets, justifying the choice of models.","marker":"[13]"},{"why":"Supplies the hyperparameter-optimization and Monte-Carlo validation methodology used to tune and verify the models.","marker":"[18]"}],"fun_headline_variants":["Self-supervised PINN nails sodium heat transfer, 8% error","87 CFD runs teach AI to predict sodium cooling within 8%","Physics-guided neural net estimates sodium Nusselt number to 8%","Transfer learning boosts sodium cooling prediction accuracy","AI balances physics and data for sodium heat sink design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the scalar physics-coefficient neuron actually couples the physics correlation into the Nusselt-number prediction, but the loss only compares that coefficient to the correlation's value and never states how the coefficient changes the output; the correlation was itself fitted to the same CFD data, so if that coupling is absent the physics-informed mechanism is empty.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised PINN nails sodium heat transfer, 8% error","87 CFD runs teach AI to predict sodium cooling within 8%","Physics-guided neural net estimates sodium Nusselt number to 8%","Transfer learning boosts sodium cooling prediction accuracy","AI balances physics and data for sodium heat sink design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2566,"prompt_tokens":1027,"completion_tokens":1539,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1455}},"tokens_in":643,"tokens_out":1539,"duration_ms":10048,"temperature":1.0,"reasoning_tokens":1455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:20:12.806247+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the central claim by ablating the physics term: train the same architecture with only the MAPE loss and compare holdout error and variance; if removing the MSE physics term leaves the 8% bound and the low variance unchanged, the physics coefficient is not doing the work attributed to it. In addition, replacing Eq. (11) with an independent correlation and re-checking the 8% bound would test whether the physics term is an independent constraint.","supporting_citations":[{"cited_title":"Enhancement of liquid sodium (na) forced convection within miniature heat sinks,","cited_arxiv_id":null,"evidence_quote":"Supplies the 87 CFD Nusselt-number data points and the empirical physics correlation (Eq. 11) that the PINN's physics loss term is built on."},{"cited_title":"Computational framework development for heat transfer studies in liquid metal-cooled small-scale heat sinks with non-circular cross-sections,","cited_arxiv_id":null,"evidence_quote":"Provides the CFD mesh and simulation framework for liquid-metal miniature heat sinks that generated the training data."},{"cited_title":"Artificial neural network modeling of nanofluid flow in a microchannel heat sink using experimental data,","cited_arxiv_id":null,"evidence_quote":"Defines the water-trained shallow NN structure and its MAPE, which transfer learning adapts to sodium."},{"cited_title":"Predictions of reynolds and nusselt numbers in turbulent convection using machine learning models,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that kernel-based and NN methods perform well on small datasets, justifying the choice of models."},{"cited_title":"Designing monte carlo simulation and an optimal machine learning to optimize and model space missions,","cited_arxiv_id":null,"evidence_quote":"Supplies the hyperparameter-optimization and Monte-Carlo validation methodology used to tune and verify the models."}],"review_version":2}