Power electronic modules PEM degrade at a continuous rate, depending on the loading applied. This implies that PEM reliability is a case-by-case study requiring frequent assessment of their health state. Finite element simulations (FES) play an important role, as they provide a mechanical description that can be used in remaining useful life RUL estimation models. They are however limited by their computational complexity, favoring simpler models that are susceptible to diverging. This work presents machine learning ML algorithms as an alternative to FES for frameworks that require frequent data updates. The state of PEMS can be described using the crack length at the contact area between the wire and the metallization of the PEMS, denoted by lc, whereas the loading corresponds to the temperature variation ΔT of the PEM for a given loading cycle. This induces thermal stress, causing the crack to grow. RUL models use mechanical quantities to describe this growth, such as stress σ, strain ϵ and displacement u. Our work consists of simulating a number of scenarios corresponding to different values of lc and ΔT, and retrieving mechanical descriptors. We created a database of 1062 scenarios, which was used to train different ML algorithms. The goal of this training is to obtain a model that can predict the output value of the FES given the input lc,ΔT even if it is different from the values seen in the training. Several algorithms were tested, and the best-performing one was incorporated in a full RUL estimation pipeline that predicts the health state of the PEMS at each cycle. The RUL pipeline is in the process of publication, and the number of scenarios is largely inferior to the total number of cycles in PEM’s lifetime. This pipeline gave accurate predictions even for the difficult tasks of interpolation and extrapolation, which demonstrates the utility of surrogate ML models for FES. ML was chosen thanks to its generalization power and low computational time, which is ×10^7 faster at inference compared to FES in our case. The full paper will also include an analysis of ML algorithms’ behaviors when trained on simulated data, to understand which types are to be preferred and why. Studying this behavior also allows us to understand which data are more important in the training process, which will optimize the data collection process. Finally, more advanced ML techniques will be tested, to efficiently transfer knowledge from simpler FES to more complex ones.