edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 4 EDFAAO (2023) 3:4-9 1537-0755/$19.00 ©ASM International® ADVANCEMENTS IN IMAGE PATTERN RECOGNITION FOR LOCK-IN THERMOGRAPHY HOTSPOT DETECTION AND CLASSIFICATION WITH SUPERVISED LEARNING Kyu Kyu Thinn1, Rui Zhen Tan2, Teh Tict Eng1, Ming Xue1 1Infineon Technologies Asia Pacific Pte. Ltd., Singapore 2Singapore Institute of Technology, Engineering cluster thinn.kyukyu@infineon.com INTRODUCTION Failure analysis (FA) plays an important role in the semiconductor manufacturing process to understand the failure mechanisms and root cause of failures. This prevents similar failures from occurring in the future and improves quality and yield. A failure is identified when there is non-conformity of IC devices to its electrical specifications according to the respective datasheet. Failure analysis often employs several steps of nondestructive techniques and destructive techniques to reveal the root cause.[1] Lock-in thermography (LIT) is a widely used nondestructive tool for detecting the failure location in ICs. LIT is effective on short and leakage failures. It involves an infrared thermal sensor to detect the surface temperature distribution. Local heating caused by leakage or short circuit can be detected by LIT, which indicates the source with a reddish signal. LIT utilizes multiple capture frames at 1 Hz acquisition rate to record temperature variations, and subsequently uses a post-processing algorithm to enhance the quality of the captured images. In a typical analysis that involves wide angle, 1X, or 5X objective, a large number of LIT images has to be processed to pinpoint critical hotspots that lead to the failure root cause. The majority of the work is carried out manually by the FA analyst. This is time-consuming and repetitive. In order to accelerate the image search process and reduce the need for human intervention, image deep learning and its classifiers, also commonly known as image pattern recognition have been demonstrated as viable solutions.[2,3] This study serves as an inspiration for the development of an intelligent tool that can assist engineers in detecting weak hotspots during LIT image processing. IMAGE PATTERN RECOGNITION FOR LIT HOTSPOT This work involves the development of an algorithm for automating image pattern recognition in LIT hotspot detection. The algorithm was created by utilizing existing annotated images obtained from the LIT analysis process. These images depict thermal signals, which are presented using a colormap and overlaid onto x-ray, scanning acoustic microscopy (SAT), or optical image backgrounds in order to visualize the location of defects (Fig. 1). It is evident that additional steps to differentiate critical thermal signals from the raw LIT images are required. A simple thresholding procedure is insufficient because of the high variance in the intensity of the thermal signal patterns and the various types of devices or packages for overlay. In general, the hotspots can appear as a single isolated signal (Fig. 2a), a diffused Fig. 1 Images of (a) raw LIT image with thermal signals (colormap) overlaid with topography background (ICs in gray), (b) image (a) overlaid with x-ray image background to visualize defect location. (b) (a)
RkJQdWJsaXNoZXIy MTMyMzg5NA==