Training Convolutional Neural Networks with Confocal Scanning Acoustic Microscopy Imaging for Power QFN Package Delamination Classification

Abstract

This study introduces a training protocol utilizing Convolutional Neural Networks (CNNs) and Confocal Scanning Acoustic Microscopy (CSAM) imaging techniques to classify Power Quad Flat No-leads (PQFN) package delamination. The investigation involves empty PQFN packages with varied substrate metallizations subjected to thermal cycling. Four delamination classes were labeled: Die-pad delamination (Class-A), Bond-pad delamination (Class-B), both Die-pad and Bond-pad delamination (Class-C), and No delamination (Class-D). Due to data imbalance, additional randomness was introduced for distribution balancing. Residual Networks (ResNet-18) based CNN model was selected for classification. Five-fold cross-validation assessed overfitting performance concerning input data size, image resolution, and batch size. The ResNet-18 prediction performance was evaluated using precision and recall metrics, with the model achieving average precision and recall scores of 0.86/1 and 0.83/1, respectively. Additionally, a comparison of delamination among different substrate metallizations was presented with Ag and NiPdAu indicating significant delamination compared to bare Cu substrate. This study pioneers the integration of CNNs with CSAM imaging for package defect detection and classification, laying the groundwork for future research to address the complex interplay of multiple failure mechanisms in functional packages.