A machine learning and finite element simulation-based void inspection for higher solder joint reliability

Abstract

We proposed a new approach for high-quality void inspection to enhance solder joint reliability. Using a small batch of samples, we developed an automatic detection algorithm for voids in the Cu-Sn solder joint. Based on ~600 experimentally obtained samples, we trained a convolutional neural network model and identified ~500 voids from ~80 samples. The obtained results indicated the voids in the solder joints were primarily located near the Cu-Sn intermetallic interface, and the averaged diameter of voids ranges from 15 μm to 25 μm. Additionally, we detected the voiding of all samples and a value below the IPC standard requirement (~15 %). However, after thermal shock cycling tests, a brittle crack was observed in a sample with 4 % voids. Based on the finite element (FE) analyses, it is found that the small interval between voids brought in a stress concentration zone under a high temperature. Meanwhile, it is found that small intervals, such as a 2.5-time-diameter of voids, weaken solder joint reliability after thermal shock cycles. A new approach, which includes deep learning-based image analysis and FE analyses, could be utilized in the solder joint quality rating to enhance reliability, particularly within autonomous driver assistance system.