Application of automated defect detection for failure analysis in scanning acoustic microscopy

Abstract

Failure analysis holds a pivotal role in semiconductor design, reliability testing and field reject investigation to ascertain the underlying causes of failure. Among the range of non-destructive imaging techniques commonly applied in failure analysis, Scanning Acoustic Microscopy (SAM) is used to discern defects, such as underfill voids and delamination between the semiconductor package interfacial layers. As semiconductor technology advances with the relentless scaling of interconnects, the human capacity to manually discern defects in acoustic images has become increasingly challenging. Various methodologies have been developed to automate SAM signal analysis for defect detection, these methods often necessitate extensive data volumes or real-time data processing. Therefore, exploring post-acquisition methods of analysis on acoustic images is in demand. The project aims to study the viability of using automation through utilizing both supervised learning and unsupervised learning approaches to identify defects from an acquired dataset of acoustic images. The acoustic image dataset was collected by scanning semiconductor packages through the SAM tool. The outcome of this study revealed promising capabilities of the approaches in identifying the “anomaly” class images. Specifically, the Area Under the Receiver Operating Characteristics Curve (AUROC) attained a value of 0.997 with the implementation of the EfficientNet architecture for the supervised learning approach and 0.814 using the vision transformer for the unsupervised learning approach.