{"id":"6561913e-08ab-4428-b601-7794a2b1ff3b","arxiv_id":"2501.16867","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A modified Stephan-Preusser correlation plus a residual-learning deep neural network predicts pool boiling HTC on microchannel surfaces with R² = 0.995 on collected literature data.","lead":"This paper proposes a new empirical correlation and a hybrid machine learning framework to predict pool boiling heat transfer coefficients on microchannel-structured surfaces. It combines a physics-based correlation with a deep neural network that learns the correlation's residuals, achieving R² of 0.995 on a compiled dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The prior correlation's exponents are fit on the whole dataset, so the PIMLAF test split is leaked into the baseline; its R²=0.995 superiority over the standalone DNN is not established.","rationale":"The reader's weakest assumption focuses on whether the nine dimensionless groups capture the physics and whether exponents are stable. That is a valid generalization concern, but the more immediate and load-bearing problem is evaluation leakage: the prior correlation is fit on the same data used to score PIMLAF. This alone can explain the large margin over the standalone DNN and would invalidate the strongest claim if confirmed. The proposed test is a single, unambiguous re-analysis: refit the correlation on training data only and re-evaluate. The paper's contribution (a hybrid residual-learning framework) is reasonable, and the issue is fixable, so the conditional verdict stands; the condition should explicitly require the leakage-free split. I partially agree with the reader: they noted the correlation is fit and evaluated on the same dataset, but they did not trace this leakage into the PIMLAF test metric, which is the sharpest form of the problem.","tokens_in":25011,"tokens_out":5627,"duration_ms":49960,"concrete_test":"Re-fit the nine correlation exponents in Eq. (2) using only the 80% training split (ideally a grouped split that keeps each source study entirely in train or test). Freeze those exponents, compute y_p on the untouched 20% test split, train the residual DNN only on training residuals, and report PIMLAF's R², MAE, RMSE alongside the standalone DNN on the same split. If PIMLAF no longer beats the DNN, or its R² drops below about 0.95, the headline performance is an artifact of in-sample fitting of the prior correlation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proposed correlation in §3.1.1 is fit to the entire 7128-point dataset and its R²=0.936 is reported on the same data. The ML pipeline in §3.2 then splits the data into 80% train / 20% test, and the PIMLAF in §3.3 uses this fitted correlation as the prior model. Consequently, for every test point the baseline prediction y_p comes from a model whose coefficients were optimized with that very test point. The residual ϵ = y - y_p is thus an in-sample residual, so the DNN only has to learn the leftover noise; the final R²=0.995 is inflated by target leakage. The comparison against the standalone DNN (trained on only 80%) is therefore unfair, and Table 6's improvement over fixed-coefficient correlations (0.55→0.936) conflates in-sample fitting with predictive skill. No grouping by source study is reported, so points from the same experiment likely appear in both train and test, adding another leakage path. Until the correlation is fit only on training data and evaluated on a held-out or leave-one-study-out split, the central claim of superior generalization is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compiles 7,128 pool-boiling data points on microchannel-structured surfaces from the literature, proposes a new empirical correlation obtained by multiplying the Stephan-Preusser correlation with nine dimensionless groups (Eq. 2), compares 18 existing correlations, trains 19 machine-learning regressors plus a deep neural network, and introduces a physics-informed machine-learning-aided framework (PIMLAF) in which a DNN learns the residuals of the proposed correlation. The authors report R² = 0.936 for the proposed correlation, R² = 0.940 for the standalone DNN, and R² = 0.995 for PIMLAF, together with SHAP-based feature importance analysis for the overall, water, and other-fluids datasets.","tokens_in":25300,"tokens_out":4011,"duration_ms":33869,"significance":"If the reported predictive performance were validated on genuinely held-out data, the work would be practically useful: a reliable HTC model for microchannel-structured surfaces is of clear engineering value, and the compilation of 7,128 points from multiple studies is a substantial contribution. The paper also provides a broad comparison of existing correlations and ML methods, and the SHAP analysis gives a physically interpretable account of feature effects. The hybrid-framework concept (correlation as prior, DNN on residuals) is a reasonable approach. However, the current evaluation protocol does not establish the central generalization claim, because the proposed correlation is fitted on the full dataset and then reused inside a train/test split, so all headline metrics are in-sample or leakage-affected numbers.","major_comments":[{"comment":"The exponents of the nine dimensionless groups in Eq. (2) are fitted to the entire 7,128-point dataset, and the reported R² = 0.936, MAE = 4.94 are evaluated on the same data. This makes the comparison against the fixed-coefficient correlations in Table 6 an in-sample fitting comparison, not a predictive comparison. The exponents should be fitted on training folds only and evaluated on held-out data (or via leave-one-study-out cross-validation) before any claim of predictive superiority is made.","section":"3.1.1 (Eq. (2))"},{"comment":"Because the prior correlation in Eq. (2) was fitted on the full dataset, the residual ε = y - ŷ_p for any point in the 20% test split is an in-sample residual: each test point already contributed to the coefficient estimates. The DNN in PIMLAF therefore learns on residuals that are artificially small, and the reported R² = 0.995, MAE = 0.907, RMSE = 2.999 do not establish that PIMLAF generalizes better than the standalone DNN, which was trained only on 80% of the data. The prior must be refitted inside each training split, or trained only on the training-study portion, before computing residuals.","section":"3.3 (PIMLAF)"},{"comment":"The manuscript does not report how the 80/20 split is stratified or grouped by source study. Since the dataset is compiled from a small number of experimental studies, near-duplicate points from the same surface and same study are likely to appear in both training and test sets, which can inflate all model metrics, including the standalone DNN. A group-wise split (e.g., leave-one-study-out), or at least a report of the study distribution across folds, is needed to assess generalization.","section":"2.4 and 3.2 (train/test split)"},{"comment":"The abstract and the conclusions state that the hybrid framework 'is able to generalize well for different datasets,' but the only evidence is the overall 80/20 split and the water/other-fluid subsets of the same dataset. These are in-sample evaluations. An external dataset or a leave-one-study-out experiment is required to support the generalization claim.","section":"Abstract and Section 4"}],"minor_comments":[{"comment":"The statement 'The percentage of data imputed in this analysis is 20%' is vague: which features had missing values, and is the imputation model fitted before or after the split? If fitted before, it is a source of leakage for all downstream models.","section":"2.4"},{"comment":"The R² values for Rohsenow (-32,809,514.6) and Pioro (-1,334,090.4) are extreme; these large negative values are likely dominated by a few outliers. Reporting MAE/RMSE on a log scale or with clipped R² would make the table more readable.","section":"Table 6"},{"comment":"Pr is listed as 'Reduced pressure (bar)' in the nomenclature, but in Eq. (2) and in several correlations Pr is also used for the Prandtl number. Please use distinct symbols (e.g., p_r and Pr).","section":"Nomenclature"},{"comment":"There are typographical errors: 'Rosenhow' should be 'Rohsenow', 'alogrithms' in the Section 3.2 heading, and the German title in reference [42] should be 'Wärmeübergang und maximale Wärmestromdichte beim Behältersieden binärer und ternärer Flüssigkeitsgemische'.","section":"General"},{"comment":"Hyperparameters are reported only for the DNN; the hyperparameter settings for the tree-based models (e.g., Extra Trees, LightGBM) after random search are not given, which limits reproducibility.","section":"3.2"},{"comment":"The paper does not state how the 7,128 data points were digitized from figures, or whether the compiled dataset will be made publicly available. Please include this information or provide the data in a repository.","section":"2.1"}],"recommendation":"major_revision","confidential_remarks":"The central weakness is a training/evaluation leakage that is fixable within the manuscript's scope. The authors should refit the correlation inside each training fold and use study-grouped splitting; after that, the comparison between PIMLAF and the standalone DNN may still be informative but must be re-run under a leakage-free protocol. The paper is otherwise clearly structured and the compiled dataset is a useful resource, so major revision is appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. This paper compiles 7128 pool boiling data points on microchannel surfaces, proposes a new empirical correlation (a modified Stephan-Preusser with nine dimensionless groups), and then wraps that correlation in a hybrid ML residual-learning framework (PIMLAF) that reports R²=0.995 on a test split. The correlation form is genuinely new, and the dataset compilation is a real asset. The paper also runs 19 ML models, a DNN, and a SHAP analysis—that's thorough.\n\nThe soft spot is the evaluation protocol, and it's load-bearing. The correlation is fit on the entire dataset: the paper says the R²=0.936 is evaluated on the whole dataset. Then the ML section splits the data 80/20 for the ML models, and PIMLAF uses that fitted correlation as its baseline. So for every test point, the baseline prediction came from a model that already saw that point during fitting. The residual the DNN learns is an in-sample residual, so the final R²=0.995 is inflated. The comparison with a standalone DNN trained only on the 80% split is therefore unfair. On top of that, the split doesn't appear to group by source study, so near-duplicate points from the same experiment likely appear in both train and test. Fixing this means fitting the correlation only on the training portion and evaluating on a held-out or leave-one-study-out split.\n\nThis matters because the paper's central claim is that the hybrid framework generalizes better than both the correlation alone and the standalone DNN. That claim is currently unsupported. The standalone DNN's 0.940 and the tree model's ~0.91 are honest numbers, and the correlation's 0.936 is a legitimate in-sample fit, but the hybrid improvement is an artifact.\n\nThe paper deserves a serious referee because the underlying material is real and the flaw is fixable. I'd send it to review but expect a major revision that redoes the evaluation without leakage. The intended audience is thermal engineers working on boiling correlations and ML-based HTC prediction. The SHAP analysis and feature engineering are also worth a look.","headline":"A new microchannel boiling correlation with a real evaluation flaw: the hybrid framework's R²=0.995 is inflated because the baseline correlation was fit on the test data.","tokens_in":25785,"tokens_out":2804,"would_cite":false,"duration_ms":24478,"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-seeded hybrid model predicts boiling heat transfer on microchannel surfaces with test-set $R^2 = 0.995$.","keywords":["Pool boiling","Microchannel structured surfaces","Heat transfer coefficient","Empirical correlation","Machine learning","Deep neural network","Physics-informed machine learning","SHAP analysis"],"falsifier":"Retrain the correlation and PIMLAF on all but one of the source studies, then predict the withheld study's HTC; if the held-out $R^2$ falls far below the reported 0.995 and 0.936, the model is memorizing study-specific offsets rather than learning transferable boiling physics.","tokens_in":102,"feed_emoji":"♨️","tokens_out":11041,"duration_ms":149366,"temperature":0.7,"pith_summary":"The paper aims to show that nucleate pool boiling heat transfer on microchannel-structured surfaces can be predicted far more accurately by combining a new empirical correlation with a deep neural network that learns the correlation's residuals than by either approach alone. Using a compiled database of 7,128 experimental points covering water, R-141b, R-123, and ethanol on copper and silicon microchannel surfaces, the proposed correlation alone reaches $R^2 = 0.936$, and the hybrid framework reaches $R^2 = 0.995$, $\\mathrm{MAE} = 0.907$ kW/m$^2$K, and $\\mathrm{RMSE} = 2.999$ kW/m$^2$K on the test set. A reliable model of this kind matters because boiling surfaces are central to cooling and energy systems, and existing correlations are accurate only over narrow operating ranges.","feed_headline":"Hybrid model hits 0.995 R-squared for boiling HTC on microchannels","feed_subtitle":"New correlation plus residual-learning network beats 19 ML models and 18 existing correlations on 7,128 data points.","key_machinery":"The load-bearing mechanism is the PIMLAF hybrid: a prior physics-based model predicts $\\hat{y}_p$, a deep neural network predicts the residual $\\hat{\\epsilon}_m$, and the final prediction is $\\hat{y}_f = \\hat{y}_p + \\hat{\\epsilon}_m$. The prior is the proposed correlation, a modified Stephan--Preusser equation whose nine dimensionless groups carry all surface-geometry and fluid-property effects through fitted exponents; the residual network (8 hidden layers, 90 neurons per layer, ELU activation, L1/L2 regularization) captures whatever systematic error remains. A feature-attribution analysis then ranks the inputs' contributions, identifying surface roughness, fin height, area augmentation factor, and groove width as the dominant parameters across datasets.","core_discovery":"The central claim is that nine dimensionless groups -- the area augmentation factor $\\lambda$, substrate-to-liquid conductivity ratio $k_w/k_l$, roughness-to-cavity-radius ratio $R_q/r_{\\mathrm{cav}}$, normalized contact angle $\\theta/90$, reduced pressure $P_r$, molecular-weight ratio $M_f/M_w$, fin aspect ratio $h_f/w_f$, groove-to-pitch ratio $w_g/p$, and hydraulic-diameter-to-pitch ratio $D_h/p$ -- can be multiplied into the Stephan--Preusser nucleate-boiling correlation with fitted exponents to describe the 7,128-point microchannel dataset with $R^2 = 0.936$ and MAE 4.94. Treating that correlation as a fixed prior and training a deep neural network to predict only the residual $\\epsilon = y - \\hat{y}_p$ gives the PIMLAF hybrid, which achieves $R^2 = 0.995$, MAE 0.907, and RMSE 2.999 on the test partition. These figures beat every one of the 19 individual ML models tested, the standalone DNN ($R^2 = 0.940$), and all 18 existing correlations evaluated, and the hybrid also reaches $R^2 = 0.992$ on a water-only subset and $R^2 = 0.997$ on other fluids.","pith_inferences":["A natural check the paper does not report is leave-one-study-out validation: if the residual network is absorbing calibration offsets specific to each source experiment, holding out an entire study should sharply reduce the hybrid's $R^2$.","The same correlation-plus-residual recipe could be carried over to other structured-surface families (pin fins, reentrant cavities, V-grooves) once datasets of comparable size exist, because the hybrid mechanism does not depend on the microchannel-specific form of the prior.","The identified dominant parameters could be turned into dimensionless design maps -- for example, HTC contours versus $h_f/w_f$ and $\\lambda$ at fixed fluid conditions -- to guide fabrication choices, although the paper stops at feature rankings."],"forward_implications":["The proposed correlation alone gives a closed-form HTC predictor for microchannel structured surfaces, improving on the best existing correlation (Stephan--Preusser) from $R^2 = 0.55$ to $R^2 = 0.936$ over the full 7,128-point dataset.","PIMLAF's test-set accuracy ($R^2 = 0.995$, MAE 0.907) is higher than every conventional ML model and the standalone DNN, showing that a physics prior plus residual learning outperforms data-only modeling on this problem.","The hybrid retains the correlation as a baseline, so predictions stay anchored to boiling physics even when the residual network is uncertain, which supports the paper's claim of better generalization to unseen datasets.","The water-only and other-fluids splits both perform well ($R^2 = 0.992$ and 0.997), indicating the framework is not relying on one fluid's behavior.","The feature-attribution analysis converts the model into design guidance: surfaces with larger area augmentation factor, taller fins, and smaller groove/fin widths and pitch should give higher HTC."],"supporting_citations":[{"why":"Supplies the Stephan--Preusser nucleate-boiling correlation that the proposed correlation modifies with the nine dimensionless groups.","marker":"[42]"},{"why":"Introduces the PIMLAF approach of combining a physics-based prior with machine-learned residuals.","marker":"[65]"},{"why":"Provides experimental pool-boiling data on various microchannel geometries used in the compiled 7,128-point dataset.","marker":"[8]"},{"why":"Provides open-microchannel pool-boiling dataset used in the compilation.","marker":"[9]"},{"why":"Supplies geometry-dependence data for open microchannel copper surfaces used in the dataset.","marker":"[26]"},{"why":"Contributes high-aspect-ratio microchannel HTC data to the compiled dataset.","marker":"[30]"},{"why":"Serves as one of the 18 existing correlations whose accuracy the proposed correlation and hybrid must exceed.","marker":"[33]"},{"why":"Serves as a widely used reduced-pressure nucleate-boiling correlation used as a comparison baseline.","marker":"[41]"}],"fun_headline_variants":["Physics-informed ML cracks boiling heat transfer on microchannels","Hybrid ML model hits 0.995 R-squared on microchannel boiling","Physics-informed hybrid beats 19 ML models on boiling HTC","Hybrid AI model achieves 0.995 R-squared for microchannel boiling","Physics-guided ML beats 18 known correlations for boiling HTC"],"cache_read_input_tokens":27904,"weakest_assumption_plain":"The load-bearing premise is that the nine chosen dimensionless parameters, with exponents fitted on the same 7,128-point dataset used to judge the correlation, fully capture how surface geometry and fluid properties control the boiling heat-transfer coefficient; if those groups are redundant, omit a governing variable, or have unstable exponents, the correlation baseline and therefore the hybrid model will not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Physics-informed ML cracks boiling heat transfer on microchannels","Hybrid ML model hits 0.995 R-squared on microchannel boiling","Physics-informed hybrid beats 19 ML models on boiling HTC","Hybrid AI model achieves 0.995 R-squared for microchannel boiling","Physics-guided ML beats 18 known correlations for boiling HTC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001053,"raw_usage":{"total_tokens":4505,"prompt_tokens":1113,"completion_tokens":3392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":3301}},"tokens_in":729,"tokens_out":3392,"duration_ms":21784,"temperature":1.0,"reasoning_tokens":3301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T10:02:57.829432+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the correlation and PIMLAF on all but one of the source studies, then predict the withheld study's HTC; if the held-out $R^2$ falls far below the reported 0.995 and 0.936, the model is memorizing study-specific offsets rather than learning transferable boiling physics.","supporting_citations":[{"cited_title":"Cooke, S","cited_arxiv_id":null,"evidence_quote":"Supplies geometry-dependence data for open microchannel copper surfaces used in the dataset."},{"cited_title":"Cooper, Heat flow rates in saturated nucleate pool boiling-a wide-ranging examination using reduced properties, in: J","cited_arxiv_id":null,"evidence_quote":"Serves as a widely used reduced-pressure nucleate-boiling correlation used as a comparison baseline."},{"cited_title":"Walunj, A","cited_arxiv_id":null,"evidence_quote":"Provides experimental pool-boiling data on various microchannel geometries used in the compiled 7,128-point dataset."},{"cited_title":"Kaniowski, R","cited_arxiv_id":null,"evidence_quote":"Provides open-microchannel pool-boiling dataset used in the compilation."}],"review_version":1}