REVIEW 3 major objections 5 minor 13 references
Deployment of ARX Models for Thermal Forecasting in Power Electronics Boards Using WBG Semiconductors
T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read 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.
desk verdict Held-out validation is genuine, but the paper sells sensor temperature as junction temperature, and that gap is load-bearing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- [§3 vs. §4; Eq. (3)] 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.
- [§2.3.1-§2.3.2, Section 3; Eq. (7)] 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.
- [Abstract, §1, §4] 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.
minor comments (5)
- [Eq. (4)] 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.
- [Eq. (8)] 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.
- [§2.3.2] 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 3] 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.
- [§2.1] 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.
Circularity Check
No load-bearing circularity: the ARX validation uses new power profiles, and the central claim does not reduce to its inputs; only a minor same-group citation and a sensor-to-junction extrapolation are noted.
full rationale
The derivation chain is linear and data-driven: Eq. (3) poses an ARX relation between power and CTN-sensor temperature; Eq. (6) estimates coefficients by least squares; Section 2.3 selects orders and the SVD regularization threshold by training fit; Section 3 then validates on two power profiles that are distinct from the training profile. The 96% and 94% validation fits are genuine out-of-sample results, not fitted parameters renamed as predictions. The near-100% training fit is explicitly presented as training-phase adjustment, not as a predictive claim. Citation [3], which shares author B. Rémy, supports the physical interpretation of ARX as an impulse-response model, but this is not the load-bearing step: ARX is a standard structure and the validation evidence stands independently. The Section 3 statement that the model estimates 'la température mesurée par les capteurs CTN' while Section 4 concludes about 'la température de jonction' is a construct-validity gap, but it is not a circular reduction because no equation identifies the sensor temperature with the junction temperature. Overall, no step is equivalent by construction to its input.
Assumptions & free parameters
free parameters (3)
- model orders (na, nb, nk) =
not reported
- SVD truncation threshold(s) =
not reported
- ARX coefficients ai, bi =
not reported
assumptions (4)
- domain assumption System is linear and time-invariant, so a SISO ARX structure with constant coefficients is adequate
- domain assumption Disturbance e[k] is white noise sharing the same denominator polynomial A as the input path
- domain assumption The NTC sensor temperature is a sufficient output to represent the thermal state of the chip
- domain assumption Physical properties and thickness variations can be ignored because identification is purely data-driven
Cite this review
Pith. "Pith review of Deployment of ARX Models for Thermal Forecasting in Power Electronics Boards Using WBG Semiconductors." pith.science (2026). https://pith.science/paper/VM4UNF72
@misc{pith2026241117748,
author = {Pith},
title = {Pith review of: Deployment of ARX Models for Thermal Forecasting in Power Electronics Boards Using WBG Semiconductors},
year = {2026},
howpublished = {\url{https://pith.science/paper/VM4UNF72}},
note = {Machine review of arXiv:2411.17748}
}
read the original abstract
Facing the thermal management challenges of Wide Bandgap (WBG) semiconductors, this study highlights the use of ARX parametric models, which provide accurate temperature predictions without requiring detailed understanding of component thickness disparities or material physical properties, relying solely on experimental measurements. These parametric models emerge as a reliable alternative to FEM simulations and conventional thermal models, significantly simplifying system identification while ensuring high result accuracy.
Reference graph
Works this paper leans on
-
[1]
Introduction L'introduction des semi-conducteurs WBG dans l'électronique de puissance marque une évolution signi ficative, posant notamment des défis de gestion thermique. En raison de leur fonctionnement à haute fréquence et sous fort courant, ces composants sont sujets à des densités de pertes importantes, provoquant une hausse critique de la températur...
-
[2]
L'Approche ARX (AutoRegressive with eXogenous inputs model) 2.1 Présentation des ARX Ljung [4] décrit que, dans le cadre théorique le plus général des modèles polynomiaux, cinq polynômes A, B, C, D et F sont identifiés. Dans le cas simplificateur d’un modèle SISO (single input single output), on a : 𝐴(𝑞−1)𝑦[𝑘] = 𝐵(𝑞−1) 𝐹(𝑞−1) 𝑞−𝑛𝑟𝑢[𝑘] + 𝐶(𝑞−1) 𝐷(𝑞−1) 𝑒[𝑘]...
work page 2024
-
[3]
Obtenir U, V, et Σ via la décomposition en valeurs singulières (SVD). Le Croisic – 18-21 juin 2024
work page 2024
-
[4]
Définir un ensemble de seuils Θ de n valeurs et initialiser le meilleur ajustement fit𝑏𝑒𝑠𝑡
-
[5]
Reconstruire la matrice Φ𝑛 à partir des valeurs singulières modifiées
Pour chaque seuil Θ𝑛 dans Θ : Modifier Σ en fixant à zéro toutes les valeurs singulières inférieures à Θ𝑛 (𝑠𝑖 𝜎𝑖 < Θ𝑛 𝑎𝑙𝑜𝑟𝑠 𝜎𝑖 = 0) . Reconstruire la matrice Φ𝑛 à partir des valeurs singulières modifiées. Évaluer l'ajustement du vecteur Φ𝑛 et mettre à jour fit𝑏𝑒𝑠𝑡 et Φ𝑛𝑒𝑤 si nécessaire (figure 2 : bloc 5 bloc 4)
-
[6]
Retourner le vecteur Φ𝑛𝑒𝑤 correspondant au meilleur ajustement trouvé (figure 2 : bloc 5 bloc 6)
-
[7]
Résultats Expérimentaux Dans notre configuration expérimentale, nous utilisons deux cartes électroniques, chacune contenant quatre puces SiC UJ4SC075011B7S disposées en deux paires parallèles, formant ainsi un bras d'onduleur. Ces cartes sont connectées pour créer un pont en H, où la distance est variable entre les puces sur les deux cartes ; on s'attend ...
work page 2024
-
[8]
Conclusion Notre étude démontre l'efficacité de l'approche ARX, combinée avec la régularisation SVD, pour la prédiction thermique dans l'électronique de puissance. Elle souligne la précision des modèles ARX dans l'estimation de la température de jonction des semi -conducteurs WBG, en se basant uniquement sur des données expérimentales. Cette méthode se di...
Show all 13 references
-
[9]
Qian et al., « Thermal Management on IGBT Power Electronic Devices and Modules », IEEE Access, vol
C. Qian et al., « Thermal Management on IGBT Power Electronic Devices and Modules », IEEE Access, vol. 6, p. 12868‑12884, 2018, doi: 10.1109/ACCESS.2018.2793300
2018
-
[10]
Birbarah et al., « Water immersion cooling of high-power density electronics », International Journal of Heat and Mass Transfer, vol
P. Birbarah et al., « Water immersion cooling of high-power density electronics », International Journal of Heat and Mass Transfer, vol. 147, p. 118918, févr. 2020, doi: 10.1016/j.ijheatmasstransfer.2019.118918
2020
-
[11]
Loussouarn, D
T. Loussouarn, D. Maillet, B. Schick, B. Rémy, D. Dan, Indirect measurement of temperature inside a furnace, ARX model identification, Journal of Physics: Conf. Series 1047 (2018)
2018
-
[12]
System identification
Ljung L. System identification. Univ.; 1995 May 29
1995
-
[13]
Improving the Signal -to-Noise-Ratio of Free Induction Decay Signals Using a New Multilinear Singular Value Decomposition -Based Filter,
H. Liu, Z. Wang, C. Zhao, J. Ge, H. Dong and Z. Liu, "Improving the Signal -to-Noise-Ratio of Free Induction Decay Signals Using a New Multilinear Singular Value Decomposition -Based Filter," in IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-11, 2021, Art...
2021
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.