Publication: Convolutional neural network-based Void detection study in underfill Encapsulation process.
No Thumbnail Available
Date
2025-09
Authors
Calvin Ling Teck Xiao
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Voiding remains a major challenge in underfill encapsulation, particularly in asymmetrical regions. Manual inspection is tedious and error-prone, while current studies mostly address symmetrical voiding. This research deploys two convolutional neural network (CNN) models with different setups to detect voids in through-scan acoustic microscope (TSAM) images of chips with epoxy underfill. Each TSAM image is analysed alongside its dispensing parameters to identify potential voiding causes via three binary classification models. Regression models are then developed for three dispensing orientations. Predictions from selected TSAM images are validated through numerical simulations replicating epoxy flow and void formation. A custom Mask RCNN ResNet101 backbone achieved a mean average precision (mAP) of 96.4%. Valve dispensing pressure emerged as the dominant cause of voiding, while
the I-type orientation yielded the best underfill quality, supported by a low mean residual of 0.001 and a prediction accuracy of 93%. Numerical simulations confirmed the CNN findings, showing voids forming in similar regions, primarily due to racing patterns leading to epoxy imbalance in the solder ball area. Additionally, a mathematical analysis using the J-integral and An and equation reveals the underfill sample is prone to creep under thermal cycling, with the critical J-integral, 𝐽-./0 being 4.0 N/mm and the critical strain rate, 𝜖-./0 of 170.5.