{"id":"626d2cac-8777-41c1-af71-d731430bade8","arxiv_id":"2502.05487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN-FCNN 'Mix Neural Network' trained on the MagNet database predicts core loss across four ferrite materials in one model, beating Steinmetz fits on the reported test set.","lead":"Using the open-source MagNet database, the authors train a combined CNN and fully connected neural network to predict magnetic core loss in ferrite materials, claiming a single model covers four materials. The work is a practical benchmark of machine learning versus Steinmetz-style equations, but some headline claims, including temperature-range generalization, outrun the provided evidence.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hybrid model weights are selected on the test set, inflating the reported improvement.","rationale":"The reader's CONDITIONAL verdict already identifies the hybrid weight search on the test set as one of the two structural issues, so my concern does not change the verdict; it reinforces the need for the stated condition. However, the reader's weakest_assumption focuses on dataset representativeness and temperature extrapolation, whereas I consider the test-set weight tuning to be the more direct threat to the paper's reported numbers. The temperature extrapolation claim is an overreach, but it can be removed without destroying the core result; the test-set tuning, if left unaddressed, leaves the hybrid model's advertised benefit unsupported. My concrete test directly settles whether the hybrid improvement is real or an artifact of test-set selection. If the test fails, the paper should either present validation-tuned results or soften the claim about the hybrid's superiority. The central single-model claim (MNN alone) is less affected, so the appropriate verdict remains CONDITIONAL rather than REJECT.","tokens_in":9966,"tokens_out":5531,"duration_ms":57975,"concrete_test":"Re-run the hybrid weight search using only the validation set (the 15% split withheld during training). Vary W1 from 0 to 1 in small increments, set W2=1-W1, and choose the pair minimizing MSE on the validation set. Then evaluate the fixed weighted predictor on the test set and report its MSE, MAPE, and R². Compare to the reported test values of MSE 334,314,144 and R²=0.997. If the validation-tuned MSE is substantially higher or R² is lower, the test-set overfitting is confirmed. Also report the validation-tuned weights and the test metrics of the individual XGBoost and MNN models for reference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-D, the authors perform a grid search over the hybrid weights W1 and W2 with the goal of minimizing MSE, obtaining W1=0.4764 and W2=0.5236, and then report the hybrid model's test-set metrics (MSE 334,314,144, MAPE 12.09%, R²=0.997) in Table VII. Because the test set is used to select the weights, the reported hybrid performance is optimistically biased: the weights are chosen to minimize test-set error, so the improvement over the individual MNN and XGBoost models may not generalize to new data. This directly affects the abstract's claim that the hybrid model's accuracy 'could continue to improve' and the reader's strongest claim that the hybrid achieves R²=0.997. The core single-model claim may still hold, but the hybrid's headline numbers are not trustworthy without a proper validation-based weight selection. This is a concrete, fixable methodological flaw rather than a fundamental theoretical issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Mix Neural Network (MNN) that combines embeddings, CNN, and fully connected layers to predict magnetic core loss from the MagNet dataset, using a single model across four ferrite materials, four temperatures, three waveforms, and frequencies from 50 to 500 kHz. The authors compare MNN with Random Forest, XGBoost, and MLP-LSTM, and with traditional Steinmetz-equation and iGSE fits, and then form a weighted hybrid of MNN and XGBoost. The central claim is that a single data-driven model can replace per-material empirical equation fits across the tested operating envelope, with accuracy far exceeding the traditional equations.","tokens_in":10144,"tokens_out":3992,"duration_ms":42540,"significance":"If the claims hold, the paper would provide a practically useful single-model alternative to per-material, per-condition empirical core-loss equations, building on the public MagNet database. The study reports clear train/validation/test splits, standard regression metrics, and a comparison of several model families, which are strengths. However, the headline hybrid accuracy is compromised by test-set-based weight selection, the temperature generalization claim is only asserted and not tested, and the comparison to traditional equations is confounded by the much richer feature set provided to the ML models. The evidence supports a narrower claim: supervised models with engineered features and deep sequence models outperform basic Steinmetz-family fits on this data subset, but not the full generalization claims in the abstract and Section III-D.","major_comments":[{"comment":"The hybrid weights W1 and W2 are selected by a grid search on the test set with the explicit objective of minimizing MSE. Consequently, the reported test-set metrics of MSE 334,314,144, MAPE 12.09%, and R²=0.997 are optimistically biased estimates of the hybrid model's generalization performance, and the statement that the hybrid 'could continue to improve' is not supported by the experiment as designed. Please select the weights using the validation set (or nested cross-validation) and report test-set metrics only for the final fixed weights.","section":"III-D, Eq. (21), Table VII"},{"comment":"The sentence claiming that the trained model 'can not only predict the core loss at these four temperature points, but also predict the entire temperature range' is unsupported by any experiment in the manuscript. All temperatures in the dataset are 25, 50, 70, and 90 °C, and no held-out temperature points are evaluated. Please either add an explicit interpolation/extrapolation experiment with temperatures not used in training or restrict the claim to the four trained temperatures.","section":"III-D, final paragraph"},{"comment":"The comparison between traditional equations and ML models is confounded by feature set. The Steinmetz and iGSE fits in Tables II and III use only frequency and flux-density information, fit per material, while the ML models receive material identity, waveform type, temperature, and the full 1024-point B waveform. The observed accuracy gap could therefore be due to the richer input features rather than the modeling method. Please add controlled experiments with matched feature sets (e.g., per-material ML models on the same inputs as SE/iGSE, or extending the equations with temperature) or explicitly frame the comparison as 'data-driven models with richer features'.","section":"III-A vs. III-B and III-D"},{"comment":"The dataset curation and split are not described in enough detail to support the generalization claim. The paper reports 12,400 samples from four materials, four temperatures, three waveforms, and frequencies from 50 to 500 kHz, but it does not state the number of samples per condition, whether the split is stratified, or whether samples from the same material/temperature/waveform condition appear in both training and test sets. Without this information, the test-set metrics do not demonstrate robustness to new material batches, temperatures, or waveforms. Please provide a detailed data sheet and split description.","section":"II-A and III-D"},{"comment":"The 33 manually extracted sequence features used for Random Forest and XGBoost are not listed or defined. Since these features form part of the input for two of the four compared models and contribute to the reported results, the experiments cannot be fully reproduced or assessed without a precise definition. Please include a table listing all hand-crafted features and their extraction formulas.","section":"III-B, feature extraction"}],"minor_comments":[{"comment":"There is a typo in the weights: the text says 'W1 = 0.4764 and W1 = 0.5236'; the second should be W2.","section":"III-D, Eq. (21)"},{"comment":"Max APE is used as a performance index in Tables II, III, and VII but is never defined. Please add its formula.","section":"II-E"},{"comment":"'Sigmod' should be 'Sigmoid'.","section":"II-C, Eq. (12)"},{"comment":"The MLP-LSTM entry lists 'patience 50' and early stopping, but the manuscript does not state what data (training or validation) is monitored for early stopping. Please clarify.","section":"III-C, Table V"},{"comment":"The y-axis of Figure 10 is described as 'error distribution', but the precise quantity (e.g., percentage error bins, MAPE per bin) is not defined in the text or caption. Please specify.","section":"III-D, Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The central data-driven claim is plausible and the use of the public MagNet dataset is a strength, but the advertised hybrid performance is inflated by test-set weight tuning and the temperature-range generalization is an untested assertion. With a validation-set-based weighting procedure, a matched-feature comparison, and held-out temperature experiments, the paper could become a solid contribution. I would also encourage the authors to compare against the recent physics-informed neural network baselines they mention (KANN, PI-MFF-CN) rather than only naming them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about applied ML in power magnetics. The central empirical claim—that a single supervised model can predict core loss across four ferrite materials, multiple temperatures, frequencies, and waveforms better than per-material Steinmetz fits—is plausible and probably correct. The MagNet database is real, the holdout split is proper, and the gap to the traditional equations is large (MAPE ~10-20% vs 30-70%). That is the paper's genuine contribution, and it is useful for engineers who currently refit coefficients per material.\n\nThe MNN itself is not a conceptual leap: it is an embedding layer for categorical features, a small CNN over the B waveform, and a FCNN. The comparison against Random Forest, XGBoost, and MLP-LSTM is fair in the sense that all get the same inputs, but it is not a controlled test of architecture. As the reader notes, the richer feature set (temperature, material embedding, waveform type, full waveform) explains much of the gap to Steinmetz; the architecture is a secondary factor.\n\nThe soft spots are real and fixable. The hybrid weight grid search is done on the test set, which inflates the reported MSE/MAPE/R² for the hybrid model in Table VII. That is a methodological slip, not a fatal one—the individual MNN and XGBoost results are still strong. The temperature generalization claim in Section III-D goes beyond the data: four discrete temperature points do not support 'predict the entire temperature range' without interpolation experiments. The feature list for RF/XGBoost is described only as '33 features extracted from sequences'; without the exact recipe, the comparison is hard to reproduce. No code is given. Minor: MLP-LSTM's poor performance might be due to poor hyperparameters, but that does not affect the main comparison.\n\nIf I were refereeing, I would ask for: (1) re-fit the hybrid weights on the validation set and re-report test metrics, (2) either test at intermediate temperatures or drop the general claim, (3) publish the full feature list and data split details, and (4) ideally release code. None of this undermines the headline direction. It deserves a serious referee and a conditional accept plan, not a desk reject. It will not change the field, but it is a clean, useful data point.","headline":"Plausible and useful applied-ML result for core-loss modeling, but the hybrid model's numbers are inflated by test-set weight tuning and the temperature generalization claim overreaches; both are fixable.","tokens_in":10664,"tokens_out":2066,"would_cite":false,"duration_ms":20933,"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":"One data-driven model can predict magnetic core loss across four ferrite materials, three waveforms, and a range of temperatures, with accuracy far above classical Steinmetz-type equations.","keywords":["magnetic core loss","core loss prediction","machine learning","deep learning","CNN-FCNN","XGBoost","Steinmetz equation","ferrite materials"],"falsifier":"Measure core loss at a temperature not used in training, for example 35 or 80 degrees Celsius, on the same four ferrite materials and waveforms, and feed the same features to the trained hybrid model; if its mean absolute percentage error rises well above the reported 12 percent or its predictions drift systematically, the claimed full-temperature-range generalization is falsified. A second check is to test on a different batch of the same material grade, since the paper only demonstrates generalization across the four grades in the training set.","tokens_in":9739,"feed_emoji":"🧲","tokens_out":11493,"duration_ms":98559,"temperature":0.7,"pith_summary":"The paper is trying to establish that a single data-driven model, trained once, can predict magnetic core loss for multiple different ferrite materials under varying temperature, frequency, and excitation waveform, with accuracy far beyond traditional empirical equations. This matters because conventional practice requires fitting separate Steinmetz-type equations for each material and operating condition, and even then prediction errors remain large. The paper reports that its proposed CNN-FCNN Mix Neural Network (MNN), and especially a weighted hybrid of MNN with XGBoost, reaches $R^2=0.997$ on held-out test samples, with mean absolute percentage error around 12 percent. If correct, this would make core-loss modeling a one-model task instead of a case-by-case fitting exercise.","feed_headline":"One trained model predicts core loss for four ferrites","feed_subtitle":"Training once on measured data covers temperatures, waveforms, and frequencies, replacing per-material curve fitting.","key_machinery":"The load-bearing machinery is the Mix Neural Network (MNN), a CNN-FCNN hybrid. Its embedding layer turns the categorical material and waveform labels into dense vectors; its two convolutional layers (16 and 32 channels, kernel size 3) read local temporal patterns from the 1024-point magnetic-flux-density waveform; and its fully connected layers concatenate the resulting features with the scalar inputs (temperature, frequency) to regress loss per unit volume. On the tree side, XGBoost handles the same task after manual feature extraction: 33 features derived from the waveform sequence plus material, waveform, temperature, and frequency. The final hybrid model is a fixed weighted average of the two, with the weights chosen by grid search to minimize MSE. The paper's argument is that giving the models the full waveform, material identity, waveform type, and temperature is what lets a single model cover conditions that classical equations either ignore or require separate fits for.","core_discovery":"On the paper's own terms, the central claim is that a single data-driven model can replace the traditional practice of fitting separate empirical equations for each magnetic material and operating condition. Using a subset of the open-source measured core-loss database—four ferrite grades (3C94, 77, N27, N87), three waveforms (sine, triangular, trapezoidal), four temperatures (25, 50, 70, 90 degrees Celsius), and frequencies from 50 to 500 kHz, with magnetic flux density given as a 1024-point sequence—the paper trains one model that handles all conditions jointly. The proposed Mix Neural Network (MNN), built from an embedding layer for categorical features, convolutional layers for the waveform sequence, and fully connected layers for regression, reaches a test mean squared error (MSE) of 593,100,352, MAPE of 20.25 percent, and $R^2=0.995$. Weighted combination of MNN with XGBoost predictions (weights 0.4764 and 0.5236) improves these numbers to MSE 334,314,144, MAPE 12.09 percent, and $R^2=0.997$. The paper further claims that because temperature enters as a continuous input feature, the trained model predicts the whole temperature range, not only the four measured points.","pith_inferences":["The accuracy gap may come more from the richer input features than from the specific neural architecture; a well-featurized tree model already matches the deep network here, so low-cost baselines should be tried before committing to deep learning on other core-loss datasets.","The full-temperature-range generalization is an interpolation claim, and it has not been demonstrated at temperatures between or outside the four training points; testing at 35 or 80 degrees Celsius would settle it.","The same input structure invites natural extensions to DC bias, core geometry, material batch variation, or higher frequencies, and whether one universal model can absorb them is an open question the paper's approach raises.","If the weights found by grid search (0.4764 for XGBoost, 0.5236 for MNN) are stable across datasets, weighted ensembling of a tree model and a deep sequence model could be a general recipe for regression tasks with mixed categorical, scalar, and sequence inputs; if the weights are dataset-specific, the reported improvement may not transfer."],"forward_implications":["A single trained model replaces the case-by-case Steinmetz fitting procedure: the paper reports accurate joint prediction for four ferrite materials in one training run.","Temperature can be treated as a continuous feature, so the model predicts not only the four measured temperatures but, by the paper's claim, the whole temperature range between them.","Non-sinusoidal waveforms are handled without waveform-specific correction equations such as iGSE, which the paper reports as having maximum errors above 370 percent for some materials.","Combining the best tree model and the best neural model through weighting improves accuracy, reaching a test $R^2$ of 0.997.","New materials or operating conditions can be added by adding data and retraining, rather than deriving a new physical equation."],"supporting_citations":[{"why":"Supplies the measured core-loss dataset (materials, temperatures, waveforms, flux-density sequences) on which every model in the paper is trained and evaluated.","marker":"[23]"},{"why":"Defines the Steinmetz equation, the classical baseline the paper fits per material and whose low accuracy motivates the machine-learning approach.","marker":"[8]"},{"why":"Introduces the iGSE correction for nonsinusoidal waveforms, the second baseline whose larger errors on triangular and trapezoidal excitations motivate the data-driven models.","marker":"[9]"},{"why":"Provides the Random Forest ensemble method used as one of the machine-learning baselines in the comparison.","marker":"[24]"},{"why":"Provides the XGBoost gradient-boosting algorithm that, together with MNN, forms the hybrid model with the best reported accuracy.","marker":"[25]"},{"why":"Supports the MLP-LSTM architecture used as the deep-learning baseline, combining MLP feature extraction with LSTM sequence processing.","marker":"[18]"}],"fun_headline_variants":["One AI model replaces per-material core loss equations","Single neural net predicts core loss for four ferrites","Hybrid MNN-XGBoost sharpens magnetic loss predictions","Core loss modeling: one model, many materials and conditions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The universal-model claim rests on the assumption that the subset of 12,400 measured samples—four ferrite materials, four temperatures, three waveforms, and 50 to 500 kHz—is representative enough for one trained model to interpolate the whole temperature range and generalize across the four materials.","fun_headline_variants_meta":{"raw":{"variants":["One AI model replaces per-material core loss equations","Single neural net predicts core loss for four ferrites","Hybrid MNN-XGBoost sharpens magnetic loss predictions","Core loss modeling: one model, many materials and conditions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2504,"prompt_tokens":1073,"completion_tokens":1431,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1365}},"tokens_in":689,"tokens_out":1431,"duration_ms":11121,"temperature":1.0,"reasoning_tokens":1365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:06:51.096641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure core loss at a temperature not used in training, for example 35 or 80 degrees Celsius, on the same four ferrite materials and waveforms, and feed the same features to the trained hybrid model; if its mean absolute percentage error rises well above the reported 12 percent or its predictions drift systematically, the claimed full-temperature-range generalization is falsified. A second check is to test on a different batch of the same material grade, since the paper only demonstrates generalization across the four grades in the training set.","supporting_citations":[{"cited_title":"Magnet: An open- source database for data-driven magnetic core loss modeling,","cited_arxiv_id":null,"evidence_quote":"Supplies the measured core-loss dataset (materials, temperatures, waveforms, flux-density sequences) on which every model in the paper is trained and evaluated."},{"cited_title":"On the law of hysteresis,","cited_arxiv_id":null,"evidence_quote":"Defines the Steinmetz equation, the classical baseline the paper fits per material and whose low accuracy motivates the machine-learning approach."},{"cited_title":"Accurate prediction of ferrite core loss with nonsinusoidal waveforms using only steinmetz parameters,","cited_arxiv_id":null,"evidence_quote":"Introduces the iGSE correction for nonsinusoidal waveforms, the second baseline whose larger errors on triangular and trapezoidal excitations motivate the data-driven models."},{"cited_title":"Xgboost: A scalable tree boosting system,","cited_arxiv_id":null,"evidence_quote":"Provides the XGBoost gradient-boosting algorithm that, together with MNN, forms the hybrid model with the best reported accuracy."},{"cited_title":"A core loss estimation method based on data-driven technology with multi- head attention mechanism,","cited_arxiv_id":null,"evidence_quote":"Supports the MLP-LSTM architecture used as the deep-learning baseline, combining MLP feature extraction with LSTM sequence processing."}],"review_version":1}