Data-driven Metamodels for Failure Analysis of Power Electronic Modules

This work presents a thorough analysis of machine learning (ML)-based metamodels to assess the reliability of power electronic modules (PEMs). Accurate lifetime prediction pipelines often require frequent updates of the mechanical state of PEMs using finite element simulations (FESs), which are computationally expensive. This compromise between complexity and precision is the motivation behind this study, which aims to harness models that are otherwise not exploitable due to the long computational time of FESs. We deem that ML models are suited for this task thanks to their fast inference speed and high predictive power. The quality of the models is thoroughly studied throughout this paper by analysing often-overlooked statistical properties that have proven to be crucial to analyse and very informative on the behaviour of the model. We applied the approach as a part of a remaining useful life (RUL) estimation pipeline for failure caused by bond-wire degradation in PEMs, but the framework is application-agnostic and can easily be adapted to different systems and degradation forms. The aim of this paper is to present a rigorous yet accessible metamodel implementation framework to appropriately exploit the strengths of ML models when applied to failure analysis.