Automated Defect Classification In Semiconductor Devices Using Deep Learning Networks

Abstract

More effective Failure Analysis (FA) technologies are required to meet the upcoming challenges in complex semiconductor devices. Because of recent advances in AI (Artificial Intelligence), we can now concentrate our efforts on developing AI-based algorithms for high precision-automated signal interpretation for failure detection in Scanning Acoustic Microscopes (SAM). Typically, flaw detection in ultrasonic data relies heavily on human expertise, and the majority of automated classifications are based on image-based decision algorithms. For defect classification, the image-based ML approach necessitates a large dataset. On signals, the traditional machine learning approach requires manual feature extraction and selection of the best features. DL approaches are commonly used to automate feature learning and classification from raw signals. This paper proposes a method for creating datasets, preprocessing signals, and semi-supervised signal training for defect classification. For performance evaluation, different DL architectures such as 1D CNN, RNN, and hybrid networks were studied. The models were trained to categorize C4 bumps in flip chips into a defect and intact classes. Even with fewer learnable, 1D-CNN with wavelet applied A-Scan as input outperforms other models with an accuracy of up to 99 percent. The model was then validated by destructive analysis on an unknown sample.