Physics-informed Machine Learning-based Methodology for Plated Through Holes Lifetime Estimation in Printed Circuit Boards

Abstract

The increasing demand for reliable electronics underscores the need for predictive tools to estimate component lifetimes and mitigate key failure risks and associated costs. This study focuses on Plated Through Holes (PTHs) in Printed Circuit Boards, which are critical to system reliability but prone to failures under standard thermal cycling due to strain from coefficient of thermal expansion mismatches. A physics-informed machine learning-based methodology is proposed, integrating data from Finite Element Method simulations and experimental data from degradation tests. Two machine learning models are combined to estimate the Remaining Useful Life of PTHs: a feedforward neural network (FFNN) able to predict the number of cycles to failure of a given structure and trained on a Design of Experiment dataset with geometric and material parameters, and a Long Short-Term Memory (LSTM) network to predict the temporal degradation trend measured by real sensors on the board. The combination of these two models allows the implementation of a Physics-informed Neural Network where the physics learned based on the FFNN is used as a physical constraint in the cost function of the LSTM to guide the prediction of the degradation.