Publication: Effect of ilu dispensing types on underfill encapsulation for void formation detection using convolutional neural network (cnn)
Loading...
Date
2024-09-01
Authors
Muhammad Taufik, Azahari
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Underfill is used to establish a mechanical and electrical connection between the ball grid array (BGA) package and the printed circuit board (PCB). Underfilling in BGA chip package is an essential process so that the reliability is ensured. There are three different types of dispensing method such as the I-type, L-type and the U-type which were commonly used in existing researches to fill the chips. Its drawback, however, are voiding which is the presence of empty space within the underfilling. The voiding defect, large in amount will cause reliability issues such as popcorning effect due to expansion of air due to heat when the chip is being used. This study aims to investigate the phenomenon of voiding in the underfilling process of large quantity BGA chips by investigating the correlation between voiding and underfilling parameters using machine learning. As manual inspection becoming less feasible to be done on large-scale manufacturing process, this study will introduce an automated method using convolutional neural network (CNN) to process the uploaded image of the chip in determining the void presence. Its percentage over the total underfilling will then be computed to identify the product can be accepted or rejected based on the
Institute for Printed Circuits (IPC) standard. Previous researches applied CNN method for variety of detection purposes, however there is no research on CNN application on underfill voiding detection. In this study, a CNN model is developed using MobileNetV2 to detect underfill voids in BGA chips and the model obtained a mean average precision of 0.533. The model detects that longer dispensing times lead to larger voids, with the I-type dispensed for 3 minutes 48 seconds having a void percentage of 1.116%, compared to 0.136% for the I-type dispensed for 2 minutes 39 seconds. Thirty unique parameters were recorded, and the results revealed that valve pressure is the primary contributor to voiding in this specific setup. Using machine learning, the best-performing model achieved an accuracy of 88.9% and an area under the receiver operating curve (AUC) of 0.722, indicating its high predictive capability.