Physics-informed Neural Network Approach for Early Degradation Trajectory Prediction of Power Semiconductor Modules

DC power cycling tests in semiconductor modules induces repetitive thermal-mechanical stresses that accumulate as fatigue over time. This paper proposes a physics-informed neural network method to reduce the reliability testing time for power semiconductor modules. The main objective of this study is to reduce the testing time while maintain a satisfactory degradation trajectory prediction accuracy using physics-informed data-driven methods. The impact of testing noise and inconsistencies from device to device is attenuated. On-state saturation voltage temperature-dependence compensation and physics based loss term regularization technique are applied in Long Short-Term Memory (LSTM) architechture, which can enhance the accuracy of degradation curve prediction under early degradation. A total of 18 IGBT devices were tested in the power cycling experiments. The proposed degradation curve prediction model can achieve an average end-of-life (EOL) prediction accuracy of 90% using approximately 40% of the early degradation testing data, which can reduce the testing time by about 60%.