Comparative study of machine learning method and response surface methodology in BGA solder joint parameter optimization

Purpose Ball grid array (BGA) package is prone to failure issues in a thermal vibration-coupled environment, such as deformation and fracture of solder joints. To predict the minimum equivalent stress of solder joints more accurately and optimize the solder joint structure, this paper aims to compare the machine learning method with response surface methodology (RSM). Design/methodology/approach This paper introduced a machine learning algorithm using Grey Wolf Optimization (GWO) Support Vector Regression (SVR) to optimize solder joint parameters. The solder joint height, spacing, solder pad diameter and thickness were the design variables, and minimizing the equivalent stress of solder joint was the optimization objective. The three dimensional finite element model of the printed circuit board assembly was verified by a modal experiment, and simulations were conducted for 25 groups of models with different parameter combinations. The simulation results were employed to train GWO-SVR to build a mathematical model and were analyzed using RSM to obtain a regression equation. Finally, GWO optimized these two methods. Findings The results show that the optimization results of GWO-SVR are closer to the simulation results than those of RSM. The minimum equivalent stress is decreased by 8.528% that of the original solution. Originality/value This study demonstrates that GWO-SVR is more precise and effective than RSM in optimizing the design of solder joints.