A Machine learning model to predict fracture of solder joints considering geometrical and environmental factors

Abstract

Predicting the fracture load and energy in solder joints is crucial for enhancing their reliability and preventing failure in electronic systems. This study introduced a novel framework by combining experimental data, machine learning (ML) models, and multi-objective optimization to predict and optimize the fracture behavior of solder joints. For this purpose, double cantilever beam (DCB) samples were fabricated and tested under displacement-control conditions and mode I crack propagation loading with a strain rate of 0.03 $s^{-1}$, incorporating environmental factors (i.e., storage temperature and humidity) and geometrical constraints (i.e., adherend thickness, adherend width, and solder thickness). The one-way analysis of variance (ANOVA) test results revealed that while all studied factors affected the fracture load of the joints, only storage temperature, humidity, and adherend thickness meaningfully influenced the samples’ fracture energy. Next, using various machine learning (ML) techniques such as artificial neural network (ANN) and random forest (RF), the solder joint’s fracture load and energy were forecasted with 86 % and 80.5 % accuracy, respectively. According to the results, the ANN and RF models predicted the joint fracture load and energy more accurately than other ML algorithms. Finally, a multi-objective optimization was employed to achieve fracture load and energy optimal values in DCB specimens by implementing the NSGA-II algorithm. This integrated approach can minimize the need for experimental testing and enable accurate prediction of solder joint fracture behavior by employing various ML models and optimization algorithms considering different working conditions.