Machine Learning Framework for Predicting Reliability of Solder Joints

This paper presents the techniques and framework to produce an explainable machine learning model for predicting the reliability of solder joints exposed to thermal cycle testing. While machine learning techniques have become prevalent for many engineering and data analysis tasks, the complexity of the model occludes valuable information which can be used to refine the data, tune model parameters, and aid in scientific discovery. By exposing the relationships within the model, users can have greater confidence in the use and employment of it. The purpose of this study is to show how useful relationships can be obtained from a traditional ”black-box” machine learning model which can be utilized to investigate the model’s ability to extract and learn fundamental physical processes.\nVarious machine learning techniques are explored with an emphasis on feature engineering, framework selection, and parameter optimization. Various model explanation methods are employed in conjunction with a generated dataset to highlight the relationships between the independent variable inputs and dependent variable output. The independent variable marginal contributions are derived from mean observations of surrogate model behavior, which can be compared to the mean observations of the machine learning model to determine relevant and exceptional behavior.\nThe application of machine learning techniques with experimental datasets certainly enables rapid evaluation of novel combinations in the problem space. While the end result is useful in itself, the ability to extract the influence of individual variable perturbations is quite challenging. The methods described herein provide this opportunity. The limitations of the model are determined by the quality and breadth of the data used to train the model parameters, in conjunction with model design specifications. While the accuracy of the determined relationships can be verified for some independent variables is possible, sparsely populated variables are less likely to generate meaningful relationships that correlate to the expected behavior of physical phenomena. This creates an opportunity for researchers to determine which data is needed to improve the model behavior in accordance with known processes. Originality/value: The ability to predict thermal fatigue life accurately is extremely valuable to the industry because it saves time and cost for product development and optimization. This ability is improved when the model can be examined critically through the methods described herein.