A Hybrid Physics and Machine Learning Digital Twin for Remaining Useful Life Prediction of SiC Power Modules in Traction Inverters

Abstract

Silicon Carbide (SiC) power modules are the most failure-prone component in automotive traction inverters, responsible for the large majority of reported inverter faults. There are two established strategies for estimating power module lifetime: a physics-based approach and a data-driven approach. Each strategy has complementary weaknesses: the physics-based approach cannot adapt to real-world degradation and is computationally heavy while pure ML approaches require large high-quality datasets, extrapolate poorly outside their training range and can produce physically unrealistic and hard-to-justify predictions. This paper proposes and evaluates a hybrid physics and ML digital twin architecture that corrects a physics-based RUL estimate by using a machine-learned residual model trained on power-cycling data from 20 SiC power modules. Input features (on-state voltage, junction temperature swing, junction temperature, thermal resistance, accumulated fatigue damage and their moving averages) are standardized and reduced via Principal Component Analysis (PCA) to a single health indicator, which together with the physics-based damage estimate feeds a residual RUL model. Six candidate regression architectures, a multilayer perceptron (MLP), Extreme Gradient Boosting (XGBoost), linear regression, ridge regression, lasso regression and a quadratic response-surface model (Box-Behnken) are compared by using 5-fold cross-validation. XGBoost and the quadratic model achieve the best trade-off between accuracy (mean R^2 ~ 0.87) and embedded implementability although the MLP achieves the highest raw accuracy (R^2 = 0.94) at higher computational cost. We further demonstrate online RUL tracking and residual-based anomaly detection under two field-data scenarios (physics-pessimistic and physics-optimistic) and discuss a path toward embedded deployment and further hybrid blending strategies.