Supervised Learning Approach in Automated Defect Detection for Failure Analysis in Scanning Acoustic Microscopy

Scanning acoustic microscopy (SAM) is a non-destructive imaging technique commonly used in failure analysis (FA) to detect defects between the semiconductor package interfacial layers. With the relentless scaling of package interconnects it is increasingly challenging for the human eye to spot defects in an acoustic image. The paper aims to study the viability of using automated defect detection through a supervised learning approach to identify defects from an acquired dataset of acoustic images. A typical dataset of acoustic images generally suffers class imbalances with more images in the “normal” class category. Our approach tackles the dataset volume increment through the usage of image patches for training. Two pre-trained convolution neural networks (CNN) are used to distinguish the “anomaly” class image patches. The Area Under the Receiver Operating Characteristics Curve (AUROC) attained a value of 0.997 with the implementation of the EfficientNet architecture.