Combining FEM and Machine Learning Algorithms for Reliability Prediction of Solder Ball Joints in BGA Packages

Abstract

This paper presents a comprehensive approach for reliability assessment of solder ball joints in Ball Grid Array (BGA) assemblies by combining finite element method (FEM) with machine learning algorithms, for instance, exploiting the power of Artificial Neural Networks (ANNs). The ANNs predictive model achieved a high score of R-square metric (0.9968 for training data, 0.9122 for test data) and low root mean square error values (43.85 for training data, 141.54 for test data), indicating robust predictive performance. In the context of the study's methodology, the ANNs model was constructed and evaluated systematically, including the analysis of loss curves and partial dependence plots (PDPs) to interpret the influence of specific solder ball geometry parameters. The results demonstrate the effectiveness of ANNs in predicting the lifetime of solder balls, implying their potential for improving reliability investigations and predictive maintenance approaches in electronics engineering.