Physics-Informed Condition Monitoring of SiC Power Modules
summary
The gist
Silicon carbide (SiC) power modules are increasingly deployed in automotive traction inverters, where reliable condition monitoring is essential to prevent in-service failures and reduce maintenance…
In short
The work addresses condition monitoring for SiC power modules using sintered packaging, where solder degradation is suppressed and wirebond liftoff dominates aging. It proposes a framework combining physics-informed features, monotonicity constraints, and heavy-tailed distribution outputs. This approach outperforms purely data-driven methods by encoding cumulative damage history to provide robust health state estimation and uncertainty quantification.
Key concepts
- Physics-Informed Feature Engineering
- This involves replacing raw sensor data with physically grounded inputs. It includes calculating cumulative damage indicators like junction temperature swings and a Miner rule accumulator. This allows the model to understand degradation history directly, making the health estimate more interpretable than purely data-driven models.
- Monotonicity Constraint
- This constraint enforces that cumulative damage indicators must be non-decreasing over time. It is implemented using gradient penalty regularization during training. This acts as a physics-guided prior, ensuring the model learns degradation in a physically realistic direction.
- Heavy-Tailed Distribution Output (Student-t)
- Instead of predicting a single health value, the model outputs parameters for a Student-t distribution. This provides calibrated uncertainty quantification, which is crucial for handling rare events like wirebond lift-offs. It captures the variance introduced by these unpredictable outliers effectively.
Terminology used across episodes
This episode discusses
- Physics-Informed Condition Monitoring of SiC Power Modules · Paper Radio
- Failure-Mechanism Transferability of Cumulative-Damage Features for Health State Estimation of SiC Power Modules · Paper Radio
The paper
Physics-Informed Condition Monitoring of SiC Power Modules · Read on arXiv
Mattia Scarpa, Evgeny Kusmenko, Francesco Toso, Mattia Bruschetta, Ruggero Carli, Simon Achatz
Department of Information Engineering, University of Padova · Infineon Technologies Dresden GmbH & Co. KG · Infineon Technologies AG
Silicon carbide (SiC) power modules are increasingly deployed in automotive traction inverters, where condition monitoring is essential to prevent in-service failures. Despite extensive qualification under AQG 324, no consolidated approach exists for in-field health state estimation: physics-of-failure lifetime models lack real-time applicability, purely data-driven architectures require large labeled datasets and generalize poorly, and physics-informed frameworks remain too demanding for embedded deployment. We address SiC MOSFET modules assembled with sintered packaging, which suppresses solder degradation and produces aging behavior distinct from previously studied devices. Instead of the smooth quasi-exponential drift of solder-based modules, the forward voltage drop V DS exhibits multi-regime profiles, with wirebond liftoff events introducing abrupt, non-monotonic perturbations. We propose a condition monitoring framework combining three elements. First, physics-informed features replace raw sensor signals with cumulative damage indicators derived from junction temperature swing, mean junction temperature and a Miner rule accumulator, encoding degradation history in an interpretable form. Second, a monotonicity constraint enforced by gradient penalty regularization embeds the expected degradation direction as a physics-guided prior. Third, a heavy-tailed output distribution replaces the point estimate, giving calibrated uncertainty robust to the out-of-distribution variance introduced by liftoff. On an industrial power cycling dataset from Infineon Technologies, several neural architectures are compared under a strict cross-validation protocol. The full configuration reduces mean absolute error by approximately 70% over purely data-driven baselines and stays stable across all folds, while remaining lightweight enough for embedded deployment.
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Physics-Informed Condition Monitoring of SiC Power Modules".
Dev: Here's the gist of the paper. Silicon carbide (SiC) power modules are increasingly deployed in automotive traction inverters, where reliable condition monitoring is essential to prevent in-service failures and reduce maintenance costs<ref:2608.08363#pg2> This work addresses condition monitoring of SiC MOSFET power modules assembled with sintered packaging technology, which prevents solder degradation and thereby produces aging behavior qualitatively distinct from previously studied devices<ref:2608.08363#pg5> The proposed framework integrates three complementary elements—physics-informed feature engineering, a monotonicity constraint, and a heavy-tailed distribution output—to address the challenges of condition monitoring in this domain<ref:2608.08363#pg2> Pruning redundant reasoning tokens The core problem addressed is the lack of a consolidated approach for in-field health state estimation of SiC devices despite extensive qualification procedures<ref:2608.08363#pg2> Existing methods are limited, ranging from physics-offailure lifetime models that lack real-time applicability to purely data-driven architectures requiring large labeled datasets<ref:2608.08363#pg2> The specific context of this study involves SiC power modules with sintered packaging technology, which suppresses solder degradation and introduces wirebond liftoff events as the dominant aging mechanism<ref:2608.08363#pg2> This scenario presents new challenges due to monotonic perturbations directly onto the forward voltage drop VDS<ref:2608.08363#pg2> How the training signal is built The proposed condition monitoring framework utilizes three distinct layers to build a robust health estimate<ref:2608.08363#pg2> First, physics-informed feature engineering replaces raw sensor signals with physically grounded inputs, including cumulative damage indicators derived from junction temperature swing and mean junction temperature, combined with a Miner rule accumulator<ref:2608.08363#pg2> Second, a monotonicity constraint is enforced via gradient penalty regularization to embed the expected degradation direction as a physics-guided prior<ref:2608.08363#pg2> Third, the model output is parameterized as a heavy-tailed distribution rather than a point estimate, providing calibrated uncertainty quantification robust to out-of-distribution variance introduced by liftoff events<ref:2608.08363#pg2> The input features include the base group of line current and relative on-state voltage drop, augmented by cumulative features such as the percycle junction-temperature swing and mean junction temperature, their integrals over time, a Miner-rule damage term, and the Miner damage of Eq.
Rosa: (4)<ref:2608.08363#pg2> Where the gains come from The framework is evaluated on an industrial power cycling dataset provided by Infineon Technologies<ref:2608.08363#pg2> The full physics-informed feature set combined with gradient penalty regularization consistently outperforms purely data-driven baselines, reducing mean absolute error by approximately 70% and maintaining stable performance across all cross-validation folds<ref:2608.08363#pg2> Specifically, the combination of cumulative features and monotonicity penalty reduces the mean absolute error (MAE) on the lifetime estimate by up to 60% relative to the baseline, with a cross-fold standard deviation of 0.019 on the health index that holds also on folds whose validation modules exhibit liftoff<ref:2608.08363#pg2> The recurrent architecture (NODE) exhibits the lowest fold-to-fold variance on every metric, consistent with its hidden state acting as an implicit memory of the trajectory<ref:2608.08363#pg2> Predictive uncertainty calibration The model output is parameterized as a Student-t distribution, where the location µn, scale σn > 0 and degrees-of-freedom νn > 0 are learned jointly through the negative log-likelihood of Eq.
Dev: (12)<ref:2608.08363#pg2> This Student-t output is then used to check whether it delivers calibrated uncertainty by counting the fraction of cycles in which the true health index rn falls inside a nominal coverage interval<ref:2608.08363#pg2> The base NODE is undercovered at both levels, reporting tighter bands than the residuals justify, whereas the cumulative variant brings the residual MAE from 7.7% to 5.0% of the unit lifetime range<ref:2608.08363#pg2> The framework successfully combines physics-informed cumulative features, a Student-t output head, and an optional monotonicity prior on top of three compact neural architectures<ref:2608.08363#pg2> The dominant lever in the comparison is the feature engineering step: encoding the cumulative thermo-mechanical history of the module is what closes the gap between the raw per-cycle measurement and a usable health estimate<ref:2608.08363#pg2> The recurrent variant offers both the best central tendency and the lowest cross-fold variance, confirming that encoding physical priors at the input level is robust across different failure modes<ref:2608.08363#pg2> References Achatz, S., Hoppe, M., Knobloch, K., Kusmenko, E., Olschewski, J., Tullmann, M., & Pfost, M.
Rosa: (2026) End-to-end snapshot models for RUL prediction of silicon carbide power modules.
Title and authors: Rosa: Let's look at what they actually propose in "Physics-Informed Condition Monitoring of SiC Power Modules." They introduce these three key elements to solve that problem we just talked about.
Dev: First, they use physics-informed feature engineering. Instead of feeding the model raw sensor signals, they build inputs based on physical damage indicators like how much the junction temperature swings and how high the mean junction temperature is over time, plus a Miner rule accumulator for damage accumulation.
Taro: That’s important because it gives the model something physically grounded to work with instead of just guessing from noise. I wonder if that kind of cumulative history helps us understand *why* a failure might be coming soon.
Rosa: Right, and then they add a monotonicity constraint enforced through gradient penalty regularization. This acts like a rule that says the degradation should always go in one direction, which helps embed the expected physics into the model's learning process.
Dev: That constraint is used to ensure that as we track time, these cumulative features actually increase or stay steady, which is a strong prior for what we expect to see during aging.
Taro: So it’s like telling the AI not to learn anything that makes sense physically wrong when it comes to how damage accumulates over time. That feels like a necessary step when you're dealing with complex failure modes.
Rosa: And finally, instead of just giving us a single number for the health state, they parameterize the model output as a heavy-tailed distribution. This is crucial because it lets them quantify uncertainty robustly around those big residuals that come from things like wirebond lift-offs.
Dev: That heavy-tailed output means they are not just predicting a single point; they’re giving us a probability distribution over the health index, which gives us a much better sense of calibration when things get weird in operation.
The paper's summary: Rosa: So, to summarize what this paper is actually doing, they’re taking the traditional way we estimate lifetime—like using Coffin-Mansontype estimates and Miner's rule for variable stress—and upgrading it significantly.
Dev: They start with a baseline that uses only the line current and relative on-state voltage drop, but then they show how adding those cumulative features and the monotonicity penalty really pushes the performance up.
Taro: The results they show are pretty compelling; specifically, combining those cumulative features with the penalty reduces their mean absolute error on lifetime estimates by up to sixty percent compared to just using the baseline approach.
Rosa: That improvement is significant because it shows that encoding that cumulative thermo-mechanical history is what really closes the gap between raw cycle measurements and a usable health estimate in practice.
Dev: They also used a recurrent architecture called NODE, which they claim has the lowest fold-to-fold variance on every metric, suggesting its hidden state acts like an implicit memory of the entire trajectory.
Taro: It’s interesting that they tested this on a dataset from Infineon Technologies—a real power cycling dataset—and showed that their cumulative features and monotonicity penalty held up even when testing modules had those lift-off events.
Rosa: So, it sounds like the main summary is that this framework successfully combines physics-informed history, a distribution output head, and an optional monotonicity prior on top of three compact neural architectures to get a robust health index.
The paper's improvements: Dev: Now let’s talk about the specific improvements they suggest over what was done before. They focus heavily on how they built the training signal and how they handled the data structure.
Rosa: The biggest improvement I see is moving from a Gaussian point estimate trained against mean squared error to outputting parameters for a Student-t distribution instead. This is done by learning the location, scale, and degrees of freedom jointly through the negative log-likelihood function.
Taro: That Student-t replaces that sharp quadratic tail with something much softer; it lets the model naturally account for those big outliers caused by events like wirebond lift-offs without breaking everything.
Dev: And they are learning that degree of freedom, or nu, per cycle, which means the reliability measure itself changes based on what's happening at that specific moment in the test. That’s a very flexible way to model uncertainty.
Rosa: Then there's the embedding of expected degradation direction through gradient penalty regularization. This isn't just tacked on; they enforce it directly onto features, ensuring that those cumulative features are non-decreasing as time goes on, which is a key physical requirement.
Taro: That monotonicity penalty seems to be what allows them to get better results when they look at the module-to-module variability, because it forces the model to respect the known physics of aging.
Dev: And for deployment suitability, they mentioned using a soft constraint for monotonicity instead of a hard construction method. This keeps the inference graph unchanged, which means we don't add any extra computational overhead when we actually deploy this on-site.
Conclusion: Rosa: So wrapping this up on "Physics-Informed Condition Monitoring of SiC Power Modules," the main implication is that encoding the cumulative thermo-mechanical history directly into the input features is what really matters for getting a usable health estimate.
Dev: The recurrent variant, NODE, provides both the best central tendency and the lowest cross-fold variance on every metric because its hidden state effectively acts as a memory of how it’s been doing throughout the test.
Taro: From my side, I think this confirms that encoding physical priors at the input level is robust even when dealing with different failure modes, which is what we need when you're trying to predict real-world behavior.
Rosa: It sounds like they've built a system that’s not just more accurate on paper but also calibrated for the messy reality of in-service power module operation.
Dev: This framework gives us a way to move toward condition monitoring that isn't just about fitting data, but about understanding the physical process happening inside the module over time.
Taro: It’s a solid step forward for health state estimation, and I’m interested to see how these cumulative features transfer when we look at different types of semiconductor modules.
More episodes
- 2608.13438-ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
- 2610.34593-Future Prediction as Training-Only Video Supervision for Pretrained VLA Policies
- 2610.35385-EmbodiedLab: Beyond Manipulation, Benchmarking Embodied Agents That Act to Know
- 2610.36378-Data-Driven Hamiltonian for Neural Optimal Control
- 2610.37275-A Heuristic for Scheduling Autonomous Vehicles at Intersections
- 2610.37288-Dynamic Coordination of Autonomous Vehicles at Intersections
- 2610.37632-Real-Time Coordination of Autonomous Vehicles at Intersections
- 2611.13779-Softness is something you also need! Why compliance helps control: information metrics of a passive walker with compliant ankle
- 2611.13411-Residual-Corrected Physics-Informed Neural Network KKL Observers with Input-to-State Stability Guarantees
- 2611.14488-Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies