A Probabilistic Bayesian Machine Learning Framework for Comprehensive Characterization of Bond Wires in IGBT Modules Under Thermomechanical Loadings

A Bayesian machine learning (ML) framework was introduced for the comprehensive characterization of bond wires within insulated gate bipolar transistor (IGBT) modules under the influence of thermomechanical loadings. The primary objective of this work was to predict two critical performance metrics, namely, equivalent plastic deformation (EPD) and the number of cycles to failure (N_f). At the core of our investigation was the dependable acquisition of training data via finite element method simulations. Based on the results, exceptional predictive accuracy was achieved, as evidenced by the impressive R-squared values of 0.962 for EPD and 0.927 for N_f, both of which are obtained from the Bayesian ML model. The high performance of the Bayesian model can be attributed to its ability to effectively capture complex relationships within the data while simultaneously being robust in handling uncertainties, rendering it suitable for situations characterized by limited datasets. Furthermore, it was revealed that the weight functions of the input parameters were significantly influenced by the values of the output targets, illustrating the distinct dependencies between each output target (EPD and N_f) and the relevant input features. These findings contribute to a deeper comprehension of the intricate interactions between input parameters and output metrics, ultimately aiding in the development of more precise and dependable models for bond wire characterization in IGBT modules.