{"id":"88625f8b-df23-4c52-aaed-8b3de54ec74b","arxiv_id":"2412.06485","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Fourier-compressed torque signal plus a Gaussian process surrogate predicts PMSM torque under geometric variations more accurately than PCA compression or no compression.","lead":"This paper builds fast stand-in models that predict the torque of an electric motor from 20 geometric design parameters, using Fourier analysis to compress the torque signal before machine learning. The best surrogate is accurate enough to replace expensive simulations in uncertainty studies, cutting cost several-fold.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Circular validation: the DFT basis (R=11, index set I) is selected from the same 2000-signal dataset later split into training and validation, so the reported MAPE does not measure generalization to unseen designs.","rationale":"The paper's stated contribution is a surrogate that predicts torque for previously unseen design configurations with best-in-class accuracy, and the strongest numerical evidence is the signal-averaged MAPE in Table 3. However, the pipeline that produces that number is circular: the DFT basis is derived from the same M=2000 dataset that is later split, so the validation set leaks into the dimension-reduction step. This is not a minor implementation detail; the choice of which 11 frequency components to keep is part of the model, and it is tuned using validation signals. Therefore the reported accuracy is optimistically biased. The reader's weakest_assumption focused on whether the fixed basis remains valid for unseen designs; the more fundamental issue is that 'unseen' is not actually true in the validation protocol. Our proposed re-evaluation with basis selection restricted to the training partition settles the question. If the DFT+GP MAPE remains lower than PCA+GP under that protocol, the claim is credible; if not, the best-in-class statement is unsupported. Given that the method is plausible and can be repaired, the appropriate verdict remains CONDITIONAL, with the additional condition that the validation be redone without leakage. We partially agree with the reader: the fixed-basis generalization concern is related, but the data leakage is the more acute and concrete problem.","tokens_in":23028,"tokens_out":11059,"duration_ms":113636,"concrete_test":"Recompute the signal-averaged MAPE for DFT+GP, PCA+GP, and no-reduction+GP at Mt=1800 with the DFT index set I and R=11 selected using only the 1800 training signals (or an independent basis-selection set), leaving the 200 validation signals completely untouched for basis construction. If the DFT+GP MAPE rises to at least the PCA+GP value, or the relative ordering changes, the claimed best-in-class advantage is an artifact of validation leakage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The DFT dimension-reduction basis is selected using the validation set. In Section 4.1, R=11 and the index set I are determined from the averaged power spectrum of M=2000 torque signals (Eq. (13)) and the reconstruction MAE (Eq. (26)). In Section 4.2, the same M=2000 dataset is partitioned into training Dt (Mt=600/1200/1800) and validation Dv=200. The validation signals therefore contribute to the averaged power spectrum eta_k used to pick I, and to the choice of R. Consequently, the MAPE values in Table 3 and Figure 5 are computed on data that influenced the reduced basis; they do not measure end-to-end generalization to a genuinely new design. This is especially damaging because the central claim is prediction for 'previously unseen design configurations' (Section 1), and because the DFT-vs-PCA gap is small (0.0029 vs 0.0030 at Mt=1800), so the reported advantage could be an artifact of this leakage. The reader's concern about the fixed basis failing outside the training distribution is related, but the more decisive problem is that the paper's validation protocol cannot detect such a failure: the 'unseen' designs were used to build the basis. The PCA comparison may suffer a similar issue if PCA components are also fitted on all 2000 samples, but the paper does not specify; either way, the absolute generalization error is not established as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven surrogate modeling framework for predicting the electromagnetic torque of a permanent magnet synchronous machine under geometric design variations. The framework combines discrete Fourier transform (DFT) based dimension reduction of torque signals with three response surface models (polynomial chaos expansion, feedforward neural network, Gaussian process). The authors report that the DFT-plus-GP combination yields the best surrogate accuracy, outperforming direct torque approximation and PCA-based reduction, and that the surrogates enable Monte Carlo uncertainty quantification at 5.5x–16.7x lower computational cost including offline expenses. The central claim is that the framework produces accurate predictions for previously unseen design configurations and reliable torque statistics.","tokens_in":23306,"tokens_out":2843,"duration_ms":28164,"significance":"If the reported results are robust, the framework is a useful contribution to surrogate-based electric machine design, with a physically motivated dimension reduction step and a systematic comparison across three regression models and three reduction schemes. The paper also provides a clear computational cost analysis and explicitly tests both uniform and Gaussian parameter distributions in the UQ studies. The main methodological strength is the use of DFT to exploit torque periodicity while retaining physical interpretability of the reduced representation. However, the validation protocol has a data-leakage issue that directly affects the reported generalization errors, and the advantage of DFT over PCA is numerically small, so the central claim requires a corrected validation before the results can be taken at face value.","major_comments":[{"comment":"The DFT basis is selected using the full dataset of M=2000 torque signals, including the Mv=200 validation samples later used in Section 4.2. Specifically, Eq. (13) averages power-spectrum contributions over all M signals to determine the index set I and the truncation R=11, and the reconstruction MAE in Eq. (26) is also computed on the same M=2000 signals. Consequently, the MAPE and SDAPE values in Table 3 and Figures 5–6 are computed on validation data that influenced the reduced basis, so they do not measure generalization to genuinely unseen design configurations as claimed in Section 1 and the Abstract. This is load-bearing because the DFT-vs-PCA gap is small (0.0029 vs 0.0030 at Mt=1800), and the reported advantage could be an artifact of leakage. The authors should either reselect the basis and R using only the training portion of the data, or use a nested cross-validation scheme, and report the resulting errors. They should also state explicitly whether the PCA components are fitted on all 2000 samples or only on the training set.","section":"Section 4.1 and Section 4.2"},{"comment":"The FNN is trained with stochastic gradient descent, but no repeated-seed or confidence-interval information is reported for the MAPE/SDAPE values. Since the FNN is the worst-performing RSM and the differences among PCE and GP are sometimes small, it is unclear whether the reported ordering of RSMs is stable. The authors should report means and standard deviations over multiple training runs with different random seeds, or at least indicate the number of runs used.","section":"Section 4.2 and Table 3"},{"comment":"In the Gaussian-distribution UQ study, the surrogate models are trained on uniformly distributed parameters but sampled under a Gaussian distribution concentrated around the nominal values. The authors acknowledge this distributional shift, but the reduced basis I was also selected from the uniform dataset. Since the Gaussian sampling region is a small subset of the training range, the fixed DFT basis may not be appropriate there. The paper should either test the reconstruction accuracy of the fixed basis within the Gaussian sampling region or discuss why the basis is expected to remain adequate; otherwise the UQ statistics errors in Table 5 may partly reflect basis mismatch rather than RSM accuracy.","section":"Section 4.3.2"}],"minor_comments":[{"comment":"In the paragraph following Eq. (3), 'Not that' should be 'Note that'.","section":"Section 2.1"},{"comment":"The caption of Figure 4 refers to 'worst-case signal reconstruction' but the worst-case signal is not defined; please clarify whether it is the signal with maximum reconstruction error.","section":"Section 4.1 and Figure 4"},{"comment":"The sentence 'In the former case, 21 principal components are kept, resulting in an MAE equal to 2.8·10−5' does not state the dataset on which this MAE is computed; please specify whether it is the same M=2000 dataset used for the DFT reconstruction study.","section":"Section 4.2"},{"comment":"The notation k(p*,P) and k(P,p*) is introduced after Eq. (23) but is used before the definitions; please define these terms in the text preceding the equation for clarity.","section":"Section 3.2.3 and Eq. (23)"}],"recommendation":"major_revision","confidential_remarks":"The data-leakage issue in the validation protocol is the main obstacle to acceptance. The paper's central claim of generalization to unseen designs is not currently supported by the reported MAPE values, since the validation samples are used to select the DFT basis. This is fixable within the manuscript's scope by redoing the validation with training-only basis selection, but it is a substantive methodological correction rather than a cosmetic one. The small DFT-over-PCA margin makes the corrected validation essential. I also note that no code or data availability statement is provided, which would be helpful for reproducibility of the numerical studies."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering study, and the DFT idea is sensible for periodic torque signals. But the validation protocol for the DFT basis is circular, and the headline advantage over PCA is small enough that the claimed best-in-class result is not established.\n\nWhat is actually new: using DFT to reduce the QoI before building response surfaces for electric machine torque, with a systematic comparison of PCE, FNN, and GP across DFT, PCA, and no reduction. The experiments are well organized and the description is clear. The UQ part is honest: they compare surrogate statistics against high-fidelity Monte Carlo, and they include offline costs in the speedup claims.\n\nThe soft spot: in section 4.1, the number of retained components R and the index set I are chosen from the full M=2000 dataset using the reconstruction MAE. Then section 4.2 splits the same 2000 into training and validation. So the validation signals were used to build the reduced basis, and the 'previously unseen' claim is not accurate for the dimension reduction step. The validation MAPE is therefore optimistic, and since the DFT-vs-PCA gap at Mt=1800 is 0.0029 vs 0.0030, that difference could easily be an artifact of the leakage. The PCA basis is likely fitted on the same 2000 as well, so the comparison may be fair in that both are affected, but the absolute generalization error is not measured. The fix is straightforward: select R and I on the training split only, or on a separate basis-selection set, and report validation on genuinely held-out designs.\n\nOther gaps are minor in comparison: no code or data shipped, no repeated-seed confidence intervals for the stochastic FNN, and the Gaussian-distribution UQ scenario already shows a 10x error increase under distribution shift. None of these are fatal.\n\nWho this is for: researchers working on surrogate models for electric machine design will find the comparison useful, but they should not take the reported numbers at face value until the leakage is fixed. It deserves a serious referee, because the method is plausible and the experimental design is mostly good, but I would not accept it as-is.","headline":"Useful engineering comparison with a real flaw: the DFT basis is selected on the full dataset, so the reported validation accuracy does not actually measure generalization to unseen designs.","tokens_in":23872,"tokens_out":2475,"would_cite":false,"duration_ms":25198,"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":"Compressing torque signals with a discrete Fourier transform and regressing the retained frequency components—best with Gaussian processes—yields the most accurate surrogate for a permanent magnet synchronous machine's torque under 20…","keywords":["discrete Fourier transform","dimension reduction","surrogate modeling","Gaussian process regression","permanent magnet synchronous machine","uncertainty quantification","polynomial chaos expansion","electric machine design"],"falsifier":"Take a design parameter combination outside the uniform training ranges in Table 1 (or a Gaussian sample near a range boundary), compute the full 120-sample torque with the high-fidelity IGA model, and compare it with the inverse DFT reconstructed from the fixed 11-component index set. If the worst-case reconstruction MAE grows well above the $9.0\\times10^{-5}$ reported for the training distribution—for instance because a formerly negligible harmonic becomes dominant—the fixed-basis assumption fails and the surrogate's accuracy ceiling is set by the truncation, not the regressor.","tokens_in":22805,"feed_emoji":"⚙️","tokens_out":8906,"duration_ms":83706,"temperature":0.7,"pith_summary":"The paper sets out to show that a cheap surrogate can replace an expensive isogeometric simulation when predicting the full torque waveform of a permanent magnet synchronous machine whose 20 geometric design parameters vary randomly. Its recipe is to compress each 120-sample torque signal with a discrete Fourier transform, keep the 11 frequency components with the largest averaged power contributions across 2000 training simulations, train a regression model from the design parameters to those components, and then invert the transform to recover the torque signal. Comparing polynomial chaos, feedforward neural networks, and Gaussian processes as regressors, the authors find that the Fourier-plus-Gaussian-process combination performs best, reaching a signal-averaged mean absolute percentage error of 0.0029 with 1800 training samples. If correct, this makes Monte Carlo uncertainty quantification of torque statistics several times cheaper, with total savings of 5.5 to 16.7 times when offline data generation and training costs are included.","feed_headline":"Fourier-then-GP surrogates predict motor torque to 0.29% error","feed_subtitle":"Combo of Fourier reduction and Gaussian processes beats direct surrogates and PCA, and cuts Monte Carlo UQ cost by up to 16x.","key_machinery":"The load-bearing object is the reduced Fourier representation. Each 120-sample torque signal $\\tau^{(m)}$ is transformed by the discrete Fourier transform into complex coefficients $c_k^{(m)}$, the per-frequency power contributions $\\eta_k^{(m)} = |c_k^{(m)}|^2$ are averaged over the $M=2000$ training signals, and the $R=11$ frequencies with the largest averaged contributions form the index set $I$ that defines the reduced quantity of interest $r\\in\\mathbb{C}^R$. A response surface $S:p\\mapsto r$, built with polynomial chaos, a neural network, or a Gaussian process, is trained on this reduced target, and prediction is $R^{-1}(S(p^*))$, the inverse DFT using only the retained modes. The fixed index set is what decouples the truncation error from the regression error and keeps the training target low-dimensional while preserving the signal's periodic content.","core_discovery":"The central claim is that periodicity makes the discrete Fourier transform a better dimension-reduction basis than principal component analysis for surrogate modeling of electric-machine torque. The authors fix a reduced basis of $R=11$ frequency components selected from the averaged power spectrum of the training set, fit response surfaces from design parameters to the complex Fourier coefficients, and reconstruct torque by inverse DFT. In their numerical comparison, DFT-based reduction outperforms direct approximation of the 120-dimensional torque signal for every regressor and training size tested, and edges out PCA for the best regressors. The best surrogate, DFT reduction with a Gaussian process response surface, reaches a signal-averaged MAPE of $0.0029$ at $M_t=1800$ and, in Monte Carlo studies, reproduces mean torque estimates to within a few ten-thousandths of the reference while cutting the cost of a $10^4$-sample study dramatically.","pith_inferences":["Beyond the paper's experiments, the averaged-power-spectrum recipe should transfer to other periodic electric-machine outputs—cogging torque, flux linkage, force ripple—where the retained harmonic set can be chosen automatically instead of from expert harmonic analysis.","A targeted robustness test would train on uniform data and then re-fit only the response surface on Gaussian-sampled data; the paper's results suggest most of the Gaussian-case error comes from distribution shift in the regressor rather than from the 11-mode DFT truncation.","An adaptive variant that re-selects the frequency index set per parameter region would be a natural extension if spectra shift with geometry; the fixed 11-mode basis is the clear point to revisit first.","For manufacturing-tolerance studies with tight distributions around nominal values, re-selecting the reduced basis on the target distribution is a low-cost safeguard, since the truncation verified on wide uniform samples may keep irrelevant modes and miss narrow-range spectral features."],"forward_implications":["With 1,800 training signals, the DFT-plus-GP surrogate predicts full torque waveforms on unseen designs with a signal-averaged MAPE of 0.0029, slightly better than the best PCA-based GP (0.0030) and far below direct no-reduction surrogates.","Monte Carlo UQ with $10^4$ samples costs about 10 seconds of online surrogate time instead of roughly 833 CPU-hours for the high-fidelity model; including offline data generation and training, total savings are 5.5 to 16.7 times depending on training set size.","For uniform parameter distributions, surrogate-based estimates of the torque mean stay within a signal-averaged APE of about $4\\times10^{-5}$ for the best GP at $M_t=1800$, with standard-deviation estimates within about 0.6%.","Training on uniform data and then sampling under a Gaussian distribution degrades statistics accuracy by roughly an order of magnitude, but the Gaussian-process surrogate still shows the smallest maximum errors and the least fluctuation.","Omitting dimension reduction altogether severely hurts all regressors, most visibly the neural network, so the reduction step itself carries much of the accuracy gain."],"supporting_citations":[{"why":"Supplies the 20-parameter PMSM benchmark geometry, material data, and the isogeometric model used to generate all torque datasets.","marker":"[21]"},{"why":"Provides the isogeometric analysis method on which the high-fidelity PMSM solver is built.","marker":"[22]"},{"why":"Gives the energy-conservation torque formula that turns the magnetic solution into the quantity of interest.","marker":"[30]"},{"why":"Provides the harmonic stator-rotor coupling used in the matrix system solved for the magnetic field.","marker":"[31]"},{"why":"Reviews combined dimension-reduction and surrogate methods and is the baseline against which the DFT-based novelty is positioned.","marker":"[23]"},{"why":"Supplies principal component analysis, the alternative dimension reduction used as a comparison baseline.","marker":"[53]"},{"why":"Provides the adaptive sparse polynomial chaos expansion via least angle regression used for the PCE response surface.","marker":"[43]"},{"why":"Supplies the Gaussian process regression framework used for the GP response surface.","marker":"[48]"}],"fun_headline_variants":["Fourier+GP surrogates hit 0.29% torque error","DFT beats PCA for motor torque surrogates","GP on Fourier basis slashes torque UQ cost 16x","Periodic torque? Fourier reduction wins for surrogates","Fourier features plus GP: best torque surrogate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that the 11 frequency components chosen from the average power spectrum of 2000 uniformly sampled training designs remain the right reduced basis for every unseen design and for the narrower Gaussian sampling distribution; if geometric variations shift the torque's spectral content into omitted frequencies, the fixed truncation adds reconstruction error the surrogate cannot correct.","fun_headline_variants_meta":{"raw":{"variants":["Fourier+GP surrogates hit 0.29% torque error","DFT beats PCA for motor torque surrogates","GP on Fourier basis slashes torque UQ cost 16x","Periodic torque? Fourier reduction wins for surrogates","Fourier features plus GP: best torque surrogate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1505,"prompt_tokens":999,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":615,"tokens_out":506,"duration_ms":5519,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:35:32.142199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a design parameter combination outside the uniform training ranges in Table 1 (or a Gaussian sample near a range boundary), compute the full 120-sample torque with the high-fidelity IGA model, and compare it with the inverse DFT reconstructed from the fixed 11-component index set. If the worst-case reconstruction MAE grows well above the $9.0\\times10^{-5}$ reported for the training distribution—for instance because a formerly negligible harmonic becomes dominant—the fixed-basis assumption fails and the surrogate's accuracy ceiling is set by the truncation, not the regressor.","supporting_citations":[{"cited_title":"‘Combined parameter and shape optimization of electric machines with isogeometric analysis’","cited_arxiv_id":null,"evidence_quote":"Supplies the 20-parameter PMSM benchmark geometry, material data, and the isogeometric model used to generate all torque datasets."},{"cited_title":"In:Computer Methods in Applied Mechanics and Engineering 194.39-41 (2005), pp","cited_arxiv_id":null,"evidence_quote":"Provides the isogeometric analysis method on which the high-fidelity PMSM solver is built."},{"cited_title":"‘On torque computation in electric machine simulation by harmonic mortar methods’","cited_arxiv_id":null,"evidence_quote":"Gives the energy-conservation torque formula that turns the magnetic solution into the quantity of interest."},{"cited_title":"‘Isogeometric analysis and harmonic stator–rotor coupling for simulating electric machines’","cited_arxiv_id":null,"evidence_quote":"Provides the harmonic stator-rotor coupling used in the matrix system solved for the magnetic field."},{"cited_title":"‘Dimensionality reduction in surrogate modeling: A review of combined methods’","cited_arxiv_id":null,"evidence_quote":"Reviews combined dimension-reduction and surrogate methods and is the baseline against which the DFT-based novelty is positioned."},{"cited_title":"‘Principal component analysis’","cited_arxiv_id":null,"evidence_quote":"Supplies principal component analysis, the alternative dimension reduction used as a comparison baseline."},{"cited_title":"‘Adaptive sparse polynomial chaos expansion based on least angle regression’","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive sparse polynomial chaos expansion via least angle regression used for the PCE response surface."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process regression framework used for the GP response surface."}],"review_version":1}