{"id":"c0a37529-2452-4da3-8b62-19f8fd7fcb0a","arxiv_id":"2411.17748","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ARX models with SVD regularization predict thermal response of SiC power boards from power input, with validation fits of 96% and 94%.","lead":"This paper tests ARX system-identification models to forecast temperatures in silicon-carbide power electronics boards from power input alone, without needing material properties. The fitted model reaches about 96 to 94 percent prediction accuracy on validation profiles, suggesting a simpler alternative to detailed thermal simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper validates ARX fits against NTC sensor temperature, but concludes about junction temperature; the unverified sensor-to-junction mapping is the load-bearing gap.","rationale":"The method is credible: ARX identification is standard, SVD regularization is a reasonable tool for ill-conditioned least squares, and validation on two held-out power profiles is a genuine predictive check. The near-100% training fit and 96%/94% validation fits do demonstrate that a linear difference model can track the NTC sensor temperature under the tested profiles. However, the paper's title, abstract, and conclusion promise junction temperature estimation for WBG semiconductors, while the experiments measure temperature at a CTN sensor placed next to the chip. The mapping between these two temperatures is not shown to be static; in real power modules the junction-to-sensor impedance involves multiple thermal time constants and power-dependent losses, so the sensor response can lag and scale differently across operating points. If that mapping is not constant, the reported fit on sensor temperature does not transfer to junction temperature. The absence of an FEM/RC baseline and the lack of error bars are also weaknesses, but the sensor-versus-junction gap is the most load-bearing because it targets the quantity the paper claims to predict. The reader flagged exactly this assumption, and I agree. The verdict should remain conditional: the evidence supports forecasting board/sensor temperature, but supporting junction-temperature claims requires the additional measurement described in the concrete test.","tokens_in":6100,"tokens_out":3005,"duration_ms":33275,"concrete_test":"Independently instrument the same H-bridge with a junction-temperature-sensitive electrical parameter (e.g., body-diode forward voltage calibrated against temperature) or an IR camera, run the same training and validation profiles, and compute the Section 3 fit from power to junction temperature. If the ARX model validates on junction temperature above 90% with bias below 5 °C, the concern is resolved; otherwise the claim must be narrowed to board/sensor temperature.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 explicitly says the ARX output is the temperature measured by CTN sensors placed beside the chip, while Section 4 claims estimation of the junction temperature of the WBG semiconductors. Nothing in the paper quantifies the relationship between these two temperatures. If the offset or the thermal lag between the NTC sensor and the junction depends on operating point (power level, PWM duty, cooling conditions), then a model identified on sensor temperature may achieve the reported 96%/94% validation fits yet fail to predict junction temperature, which is the quantity that matters for reliability and for comparison with FEM/RC models. The ARX coefficients are fit directly to sensor T in Eqs. (3)-(6); no correction term or uncertainty bound is added. Thus the central advertised capability is not established by the data shown.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ARX modeling approach, regularized by singular value decomposition, to predict thermal response on power electronics boards with WBG semiconductors from experimental power and temperature measurements. A model is identified on one power profile and validated on two held-out profiles, with reported fits of near 100% on training and 96% and 94% on the two validation profiles. The authors conclude that the ARX model is a reliable alternative to FEM simulations and conventional thermal models without requiring material property or thickness information.","tokens_in":6159,"tokens_out":5741,"duration_ms":55616,"significance":"The identification procedure is standard and the held-out validation on two profiles is a genuine positive: the reported 96% and 94% fits are predictions on data not used for fitting, not just interpolation. If the claims were limited to the measured NTC sensor temperature, the paper would offer a simple, reproducible, data-driven estimator for board-level thermal monitoring. The significance is currently limited by the unquantified relationship between the NTC sensor temperature and the junction temperature claimed in the conclusions, by the absence of any comparison with FEM or RC models, and by the small validation set with no uncertainty quantification.","major_comments":[{"comment":"The output y[k] in Eq. (3) is the temperature measured by the CTN/NTC sensors placed beside the chip (Section 3), but Section 4 concludes that the approach estimates the junction temperature of the WBG semiconductors. The relationship between sensor temperature and junction temperature is never quantified. If the offset or dynamic lag between sensor and junction varies with operating point (power level, PWM duty, cooling conditions), a model fitted to sensor temperature can achieve the reported 96% and 94% validation fits yet fail to predict the junction temperature, which is the quantity relevant for reliability and for comparison with FEM/RC models. Please either add a measurement or estimation of junction temperature, or restrict the claims to NTC sensor temperature and explain how the sensor measurement can be used as a proxy for the junction.","section":"§3 vs. §4; Eq. (3)"},{"comment":"The near-100% training fit is an expected artifact of selecting the ARX orders na, nb, nk and the SVD threshold by maximizing the fit on the training set. The only generalization evidence is two validation profiles, reported as global fit values in Eq. (7) with no error bars, no residual analysis, no confidence intervals, and no sensitivity study with respect to model order or SVD threshold. The fit metric in Eq. (7) is a normalized norm ratio that may be dominated by the slow thermal baseline rather than by the transient behavior of interest. Please report residuals, uncertainty bounds, and results on a larger set of validation profiles, and clarify how the AIC criterion mentioned in §2.3.1 was used or why it was not used.","section":"§2.3.1-§2.3.2, Section 3; Eq. (7)"},{"comment":"The abstract and conclusion claim that ARX models are a reliable alternative to FEM simulations and conventional thermal models, but the manuscript contains no comparison of the ARX predictions with FEM, RC-ladder, or any other thermal model on the same experimental data. The paper can show that ARX fits the measured sensor temperature, but it does not presently establish that it is an 'alternative' to FEM/RC in the claimed sense. Please add a baseline comparison on the same validation profiles, or explicitly soften the claim to a data-driven complement rather than an alternative.","section":"Abstract, §1, §4"}],"minor_comments":[{"comment":"The exogenous regressor block in Eq. (4) appears shifted by one sample relative to Eq. (2): for the first row, the b1 regressor should be P[1-nk], not P[-nk]. In addition, the last entry of the parameter vector is written as -b_nb, whereas Eq. (3) enters b_i with a positive sign. Please correct the indices and signs so that the regression matrix is consistent with the model equation.","section":"Eq. (4)"},{"comment":"Eq. (8) states Ω = U Σ V^T, but singular value decomposition applies to Φ, not to the parameter vector Ω. This is inconsistent with Eq. (6) and with the surrounding text, which correctly describes the decomposition of Φ. Please revise the notation.","section":"Eq. (8)"},{"comment":"The SVD algorithm text says 'évaluer l'ajustement du vecteur Φ_n' and returns Φ_new, but what is actually selected is the estimated parameter vector Ω obtained from the truncated regressor matrix. Please rephrase the procedure so that the object being evaluated and returned is unambiguous.","section":"§2.3.2"},{"comment":"The claimed 20 °C hot-spot difference between the two board layouts is reported without measurement uncertainty or calibration information for the CTN sensors; please state sensor accuracy and the thermal coupling between the sensor and the chip.","section":"Section 3"},{"comment":"Table 1 and the references [3]-[5] would benefit from more precise in-text citations; for example, the claim that ARX models can be converted to pure convolutive models should cite [3] at the point of the claim rather than only in the surrounding paragraph.","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a compact conference contribution with genuine held-out validation, and the identification methodology is sound. The decisive issue is the unquantified gap between the NTC sensor temperature used as the model output and the junction temperature claimed in the conclusions. If the authors can measure or estimate junction temperature, or carefully reframe the claims to sensor temperature, the work would be within scope. The absence of any FEM/RC comparison should also be addressed. I would not reject the paper, but the advertised contribution is currently broader than what the data and analysis support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean application of standard ARX identification with SVD regularization to temperature forecasting on a SiC power board. The genuinely new part is the specific experimental setup and the two held-out validation profiles, with 96% and 94% fits. That is a real predictive check, not a training-fit claim. The authors also deserve credit for describing the iterative selection procedure and for using a second power profile that stresses different operating points.\n\nThe math is textbook Ljung-style least squares, and the SVD filtering is standard. No new theory, and the paper doesn't claim any.\n\nThe soft spots are in the claims, not the arithmetic. First, the output variable is the temperature measured by NTC/CTN sensors placed next to the chips, but the abstract and conclusion talk about junction temperature. Nothing in the paper quantifies the offset or dynamic lag between the sensor reading and the actual junction temperature. If that relation changes with power level, PWM duty, or cooling, a model tuned to sensor temperature can hit 96% validation and still mispredict the junction. That is a load-bearing gap, and the stress-test note is right to call it out.\n\nSecond, the abstract's 'reliable alternative to FEM and RC models' is not backed by any baseline comparison. You get numbers for the ARX fits, but no FEM or RC runs on the same profiles, so 'alternative' is just a hope.\n\nThird, the near-100% training fit is inflated by model-order and SVD-threshold selection on the training data. The authors acknowledge the overfitting risk and do validate on held-out profiles, so this is a minor issue, but it means the training fit shouldn't be reported as a headline number.\n\nAlso minor: only two validation profiles, no error bars or residual analysis, no code or data. That limits independent checking but doesn't sink the method.\n\nWho is this for? Practitioners in power electronics thermal monitoring who want a data-driven observer and don't need physical insight. A serious referee should engage with it, mainly because the sensor-to-junction gap is fixable with a short calibration section. As it stands, the paper should be conditionally accepted at most, with the junction-temperature language toned down or supported.","headline":"Held-out validation is genuine, but the paper sells sensor temperature as junction temperature, and that gap is load-bearing.","tokens_in":6802,"tokens_out":2521,"would_cite":false,"duration_ms":23728,"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":"The paper's central claim is that ARX models trained only on measured power and NTC temperature can predict WBG power-board temperatures to within a few percent on unseen operating profiles.","keywords":["ARX models","system identification","thermal modeling","wide-bandgap semiconductors","power electronics","SVD regularization","temperature prediction","SiC modules"],"falsifier":"Put a directly measured junction-temperature signal, for example from the chip's on-state voltage drop or infrared thermography, next to the model's sensor-temperature prediction across power profiles with fast transients and different steady-state levels. If the sensor-to-junction offset or lag changes with operating point or board layout, the ARX model trained on sensor temperature will not track junction temperature with the same 94-96% fidelity.","tokens_in":5858,"feed_emoji":"🔥","tokens_out":6082,"duration_ms":54200,"temperature":0.7,"pith_summary":"The paper argues that a simple parametric ARX model, identified by ordinary least squares from experimental power and temperature measurements alone, can forecast chip temperatures on wide-bandgap semiconductor power boards accurately enough to be a practical alternative to finite-element and RC-ladder thermal models. On a SiC H-bridge test bench with two board layouts, the training fit reaches near 100%, and model predictions on two previously unseen power profiles reach 96% and 94%. The practical payoff is that thermal monitoring can skip detailed material property data and thickness information, which are hard to obtain for thin WBG die and substrates. A careful reader will note that the predicted quantity is the temperature measured by NTC sensors placed beside the chips, while the conclusion speaks of junction temperature.","feed_headline":"Simple ARX models forecast board heat to 94-96%","feed_subtitle":"Power and temperature data alone train WBG board thermal models, skipping material properties.","key_machinery":"The central object is the ARX (AutoRegressive with eXogenous inputs) discrete-time model. It is a special case of the general polynomial model family in which the output $y[k]$ is a linear combination of its own past values and past values of the input, plus a noise term sharing the same denominator polynomial. The paper's version sets output to board temperature and input to dissipated power, then recasts the recurrence as the linear regression $T = \\Phi\\Omega + e$, where the columns of $\\Phi$ are lagged temperatures and lagged powers. Ordinary least squares gives $\\hat{\\Omega}$; a singular-value-decomposition regularization step, implemented by zeroing singular values below a scanned threshold, stabilizes the inversion when the regression matrix is ill-conditioned under measurement noise.","core_discovery":"In the paper's own terms, the discovery is that a discrete-time ARX model of the form $T[k] = -\\sum_{i=1}^{n_a} a_i T[k-i] + \\sum_{i=1}^{n_b} b_i P[k-i-n_k+1] + e[k]$, with power $P$ as input and temperature $T$ as output, captures the thermal dynamics of a WBG power board without any knowledge of layer thicknesses or material thermal properties. The parameters are identified by least squares on a regression matrix built from measured power and temperature, after subtracting ambient temperature; singular-value-decomposition regularization truncates small singular values to damp noise before inversion. A grid search over autoregressive order $n_a$, exogenous order $n_b$, and input delay $n_k$ selects the triplet with best fit, and the retained model is validated on new power profiles. Reported fits are near 100% on training data, 96% on a multi-operating-point validation profile, and 94% on a second profile with longer steady-state segments.","pith_inferences":["One unstated consequence is that the ARX model, being linear and time-invariant, is best suited to operating regions where the thermal response is approximately linear; strongly temperature-dependent material properties may require local model scheduling or a nonlinear variant.","The method could be turned into a health-monitoring tool: if the identified ARX coefficients drift with ageing or solder degradation, the change in the impulse response may be a diagnostic signature, but the paper does not test this.","A direct comparison of SVD-regularized predictions against unregularized least squares on the same profiles would isolate how much of the accuracy is due to regularization rather than to the ARX structure itself.","The paper's sensor-level validation suggests a concrete next experiment: recalibrate the model on measured junction temperature and compare the validation fits, which would test whether the junction-temperature claim survives the sensor-to-chip transfer."],"forward_implications":["Thermal models for a new WBG board can be obtained from a single power-excitation experiment plus a temperature sensor, without extracting material properties or measuring layer thicknesses.","The identified ARX model can be used for online temperature estimation and thermal protection during operation, since it is cheap to evaluate in real time.","Board layout differences, such as whether chips are close together or far apart, show up in the measured data and are absorbed by the identified model, as seen in the roughly 20 °C hot-spot difference between the two boards.","The same identification loop with SVD regularization can be rerun for each new profile, supporting an iterative, automated model-selection workflow.","Adding extra measurable inputs, such as baseplate temperature, can be folded into the ARX structure if one wants refined predictions."],"supporting_citations":[{"why":"Documents the thermal-management challenge of power semiconductor devices that motivates junction-temperature estimation.","marker":"[1]"},{"why":"Provides the high-power-density electronics cooling context that supports the thermal stress motivation.","marker":"[2]"},{"why":"Shows ARX models can be converted to pure convolutive models, the property the authors cite for choosing ARX over other polynomial structures.","marker":"[3]"},{"why":"Supplies the system-identification framework in which the ARX polynomial structure and least-squares identification are defined.","marker":"[4]"},{"why":"Supplies the singular-value-decomposition-based filtering and regularization technique used to stabilize the regression inversion.","marker":"[5]"}],"fun_headline_variants":["ARX models forecast WBG board heat to 96% without material data","Power and temperature only: ARX predicts board temps to 96%","ARX fills thermal model gap for WBG boards using just measurements","Skip FEM: ARX models hit 96% accuracy on WBG board heat","Data-driven ARX predicts WBG board temperatures up to 96%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the temperature recorded by NTC sensors placed beside the chips is a faithful stand-in for the SiC junction temperature; if the offset or dynamic lag between sensor and junction varies with operating point, the high validation fits on sensor temperature do not automatically transfer to the junction temperature the conclusions invoke.","fun_headline_variants_meta":{"raw":{"variants":["ARX models forecast WBG board heat to 96% without material data","Power and temperature only: ARX predicts board temps to 96%","ARX fills thermal model gap for WBG boards using just measurements","Skip FEM: ARX models hit 96% accuracy on WBG board heat","Data-driven ARX predicts WBG board temperatures up to 96%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3163,"prompt_tokens":809,"completion_tokens":2354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":2256}},"tokens_in":425,"tokens_out":2354,"duration_ms":15515,"temperature":1.0,"reasoning_tokens":2256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:08:51.684551+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Put a directly measured junction-temperature signal, for example from the chip's on-state voltage drop or infrared thermography, next to the model's sensor-temperature prediction across power profiles with fast transients and different steady-state levels. If the sensor-to-junction offset or lag changes with operating point or board layout, the ARX model trained on sensor temperature will not track junction temperature with the same 94-96% fidelity.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the thermal-management challenge of power semiconductor devices that motivates junction-temperature estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the high-power-density electronics cooling context that supports the thermal stress motivation."},{"cited_title":"Le Croisic – 18-21 juin 2024","cited_arxiv_id":null,"evidence_quote":"Shows ARX models can be converted to pure convolutive models, the property the authors cite for choosing ARX over other polynomial structures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the system-identification framework in which the ARX polynomial structure and least-squares identification are defined."},{"cited_title":" Reconstruire la matrice Φ𝑛 à partir des valeurs singulières modifiées","cited_arxiv_id":null,"evidence_quote":"Supplies the singular-value-decomposition-based filtering and regularization technique used to stabilize the regression inversion."}],"review_version":1}