Physics-informed Machine Learning for Predicting Fatigue Damage of Wire Bonds in Power Electronic Modules

This paper details a novel physics-informed data-driven approach for developing computationally fast metamodels for predicting fatigue damage and its spatial distribution at common failure sites of power electronic components. The proposed metamodels aim to serve the end-users of these power components by allowing an informative model-based assessment of the thermal fatigue damage in the assembly materials due to different application-specific, qualification and user-defined load conditions, removing current requirements for comprehensive device characterisations and deploying complex finite element (FE) models. The proposed methodology is demonstrated with two different metamodel structures, a multi-quadratic function, and a neural network, for the problem of predicting the thermal fatigue damage due to temperature cycling loads in the wire bonds of an IGBT power electronic module (PEM). The results confirmed that the proposed approach and the modelling technology can offer FE-matching accuracy and capability to map highly nonlinear spatial distributions of the damage parameter over local sub-domains associated with material fatigue degradation and failure due to material/interfacial cracking.