Abstracts by Session

Session E: AI Applications for Failure Analysis
Wed   11:00 – 11:40 Anna Safont AI in Failure Analysis: Applications and Benefits

Failure analysis in semiconductors is a tedious and knowledge-intensive task. Even though highly skilled engineers perform this task, their knowledge about the considered samples might be insufficient, thus, increasing the time and costs of analysis jobs. Modern Artificial Intelligence (AI) methods can simplify the work of engineers by providing them with additional information about the current job. This tutorial will provide an overview of both knowledge-based and machine learning methods that can be applied to achieve the goals mentioned above. In particular, we focus on various applications of artificial neural networks (ANNs), helping to locate parts of an image with possible failures, automatically search databases for similar jobs performed in the past, or predict tasks that might lead to faster failure localization. We will cover general ideas behind ANNs and discuss modern architectures applied to solve problems relevant to failure analysis, such as image and text classification. The discussion will be illustrated by selected case studies demonstrating the discussed AI methods in practical applications.

Wed  11:40 – 12:00 Jeff Gelb High Speed 3D X-ray Imaging of Defects at Submicron Resolution in Large Electronic Components

3D X-ray techniques have grown tremendously in recent years and are now essential tools for electronic device FA. The long acquisition times required for high resolution, however, along with the presence of CT artifacts on larger samples, have limited the application space and introduced challenging issues for FA engineers. While 3D X-ray technologies are exciting for the future of non-destructive FA, more innovations are needed from the CT industry to support adoption.

In our work, we have developed a fast & flexible 3D X-ray system, with resolution to sub-0.5 um and integration times of only 5-10 minutes. The virtual slices produced are nearly artifact-free, well-suited for processing and data analytics. We have also established a workflow for automated data processing, yielding custom reports for various parameters (e.g.: wets vs. non-wets, cracks/voids, warpage, etc.). This technique is independent of sample size, allowing FA engineers to quickly locate defects in a variety of parts, from PCBs to full wafers. With the results presented here, we demonstrate a powerful new approach to NDT and also suggest a potential path forward proactive (e.g., in-line or near-line) solutions.

Wed  12:00 – 12:20 Eckhard Sperschneider Automatic-X-Ray inspection in combination with AI analysing methods

Automatic X-Ray Inspection in combination with AI-Analysing methods: How can machine-learning techniques improve the performance of critical inspection tasks in electronic\’s and component manufacturing?

This will be demonstrated on 2 typical production use cases:

  • Void inspection in multi-layer power-hybrid components
  • Battery inspection with Anode-Cathode distance measurements

X-RAY images generated from a Standard 2D-AXI system and with a CT-AXI setup – both using a photon-counting detector with 16/32bit imaging capability, where processed with state of the art ML techniques.Demonstration of Inspection results comparison to existing inspection techniques.

Wed  12:20 – 12:40 Sebastian Brand Implementation of Machine Learning Approaches for Automating Data Interpretation in Failure Analysis: Signal Analysis in Acoustic Microscopy for Defect Localization and Material Characterization
The increasing level of assistance and automation not only in the mobility sector demands a high reliability of increasingly complex microelectronic components and systems. To ensure the required performance and the necessary safety evolving technologies need to be understood in their behaviour down to the formation and propagation of defects and the interaction of the involved materials under operational conditions. In this respect capable testing and inspection techniques are required for analysis, but also for screening during production processes to maintain high quality levels, which is connected to large quantities of data. The paper describes the application of deep learning techniques for enhancing the analysis of acoustic time-domain signals for decision making and material characterization. Microelectronic systems commonly contain complex structured architectures. Unfortunately, this complexity and the correspondingly large number of material interfaces and their small dimensions result in the occurrence of multiple overlapping pulses in the received signals, challenging accurate interpretation. However, the obtained signals precisely represent the interaction of the acoustic wave with the samples. Machine learning is employed here to extract characteristic signal features to non-destructively obtain information which is related to physical properties of the sample under investigation and consequently detect and localize defects.
Wed  12:40 – 13:00 Matias Oscar Volman Stern Deep Learning-assisted microscopic acquisition of microelectronic components with integrated region-of-interest feature extraction

It is often necessary to acquire several regions of a sample, either with different objectives or to observe changes after some work has been applied to it. However, taking a complete image of the sample can be time consuming when only a small portion of it needs to be studied, as well as the complexity and appearance of features in 3D can vary. Therefore, we present a nearly fully automated workflow for the analysis of a specific use case, bond pad detection for cratering tests.

The work is grouped into an automated optical microscopic job (ZEISS ZEN core). Starting with a large image acquisition, followed by a deep learning model to detect the bond pads and finished with an image analysis to extract their coordinates. These are then used for future acquisitions of the modified sample to ensure high repeatability and accuracy in detecting the relevant positions.

A U-Net model was trained, with EfficientNetB3 as encoder, with RGB patches of 1024 x 1024 pixels and the corresponding binary masks. Sigmoid and Dice-Focal were used as the activation and compound loss functions respectively and the evaluation metrics were the mean IoU and mean F1-Score, both achieving more than 99%.

Wed  13:00 – 13:20 Khaled Alsaih Deep Learning Denoising of SEM Images Based on Denoiseg Trained with Synthetic Images
Microscopy image investigation mostly requires the segmentation of targeted elements, but training data is typically tight and challenging to obtain due to the unavailability of ground truth images. Here we trained a deep learning end-to-end model named DenoiSeg on synthetic data generated using physically-based noise models. Denoiseg is an extension of Noise2Void. Unlike other architectures, DenoiSeg takes into account the segmentation and denoising information which are trained jointly in an encoder-decoder style. We generated three synthetic datasets which mimicked the SEM real images, called 1kV-10pA, 2kV-10pA, and 2kV-38pA. Each dataset contains 3000 images and during the training, the augmentation is performed. Data trained is validated using 4k fold cross-validation. The performance evaluation of the three trained networks is performed using two metrics, namely, Peak signal-to-noise rate (PSNR) and Similarity index measurement (SSIM). We obtained first results: the improved PSNR for the three datasets are 13.97, 13.68, and 13.78, respectively ; the Improved SSIM for the datasets are 0.48, 0.46, 0.5. To sum up, these results are promising using the synthetic data.