Minor flaws are becoming extremely relevant as the complexity of the semiconductor package evolves. Scanning acoustic microscopy is one analytical tool for detecting flaws in such a complex package. Minor changes in the reflected signal that could indicate a fault can be lost during image reconstruction, despite the high sensitivity. Because of recent AI (Artificial Intelligence) advancements, more emphasis is being placed on developing AI-based algorithms for high precision-automated signal interpretation for failure detection. This paper presents a new deep learning model for classifying ultrasound signals based on the ResNet architecture with 1D convolution layers. The developed model was validated on two test case scenarios. One use case was the detection of voids in the die attach, the other the detection of cracks below bumps in Flip-chip samples. The model was trained to classify signals into different classes. Even with a small dataset, experiment results confirmed that the model predicts with a 98 percent accuracy. This type of signal-based model could be extremely useful in situations where obtaining large amounts of labeled image data is difficult. Through this work we propose an intelligent signal classification methodology to automate high volume failure analysis in semiconductor devices.

Failure Analysis (FA) is a complex activity that requires careful and complete documentation of all findings and conclusions to preserve knowledge acquired by engineers in this process. Modern FA systems store this data in text or image formats and organize it in databases, file shares, wikis, or other human-readable forms. Given a large volume of generated FA data, navigating it or searching for particular information is hard since machines cannot process the stored knowledge automatically and require much interaction with experts. In this paper, we investigate applications of modern Natural Language Processing (NLP) approaches to the classification of FA texts with respect to electrical and/or physical failures they describe. In particular, we study the efficiency of pretrained Language Models (LM) in the semiconductors domain for text classification with deep neural networks. Evaluation results of LMs show that their vocabulary is not suitable for FA applications, and the best classification accuracy of appr. 60% and 70% for physical and electrical failures, respectively, can only be reached with fine-tuning techniques.

The development of intelligent assistants helping Failure Analysis (FA) engineers in their daily work is essential to any digitalization strategy. In particular, these systems must solve various computer vision or natural language processing problems to select the most critical information from heterogeneous data, like images or texts, and present it to the users. Modern artificial intelligence (AI) techniques approach these tasks with machine learning (ML) methods. The latter, however, require large volumes of training data to create models to solve the required problems. In most cases, enterprise clouds store vast volumes of data captured while applying various FA methods. Nevertheless, this data is useless for ML training algorithms since it is stored in forms that can only be interpreted by highly-trained specialists. In this paper, we present an approach to embedding an annotation process in the everyday routines of FA engineers. Its services can easily be embedded in existing software solutions to (i) capture and store the semantics of each data piece in machine-readable form, as well as (ii) provide predictions of ML models trained on previously annotated data to simplify the annotation task. Preliminary experiments of the built prototype show that the extension of an image editor used by FA engineers with the services provided by the infrastructure can significantly simplify and speed up the annotation process.

This paper introduces the use of machine learning models in the characterization of bitmap fail patterns occurring on SRAM to identify FEOL/MEOL layers defectivity distribution. The results of bitmap patterns with test conditions are used for fault analysis post-processing and manufacturing yield improvement methodologies. Several machine learning models were built for prediction of the FEOL/MEOL layer defects based on hundreds of bitmap physical failure analysis results. A model utilizing a multilayer perceptron (MLP) architecture with backpropagation of error were optimized and it can be easily applied to volume products with millions of bitmap test results with >80% accuracy. It is the first time we are able to investigate the FEOL/MEOL defects density quantitatively through an automatic diagnosis tool.

Object localization is an essential step in image-based hardware assurance applications to navigate the view to the target location. Existing localization methods are well-developed for applications in many other research fields; however, limited study has been conducted to explore an accurate yet efficient solution in hardware assurance domain. To this end, this paper discusses the challenges of leveraging existing object localization methods from three aspects using the example scenario of IC Trojan detection and proposes a novel knowledge-based object localization method. The proposed method is inspired by the 2D string search algorithm; it also couples a mask window to preserve target topology, which enables multi-target localization. Evaluations are conducted on 61 test cases from five images of three node-technologies. The results validate the accuracy, time-efficiency, and the generalizability of the proposed method of locating multi-target from SEM images for hardware assurance applications.