A RESOURCE FOR TECHNICAL INFORMATION AND INDUSTRY DEVELOPMENTS AUGUST 2023 | VOLUME 25 | ISSUE 3 ELECTRONIC DEVICE FAILURE ANALYSIS edfas.org ADVANCES IN PATTERN RECOGNITION FOR LIT HOTSPOT DETECTION FA OF PHOTONIC INTEGRATED CIRCUITS FLUORESCENCE LIFETIME IMAGING TO IDENTIFY PLASTICS FOUR-DIMENSIONAL STEM: IMAGING, STRAIN, AND DEFECTS 4 23 12 31
A RESOURCE FOR TECHNICAL INFORMATION AND INDUSTRY DEVELOPMENTS AUGUST 2023 | VOLUME 25 | ISSUE 3 ELECTRONIC DEVICE FAILURE ANALYSIS edfas.org ADVANCES IN PATTERN RECOGNITION FOR LIT HOTSPOT DETECTION FA OF PHOTONIC INTEGRATED CIRCUITS FLUORESCENCE LIFETIME IMAGING TO IDENTIFY PLASTICS FOUR-DIMENSIONAL STEM: IMAGING, STRAIN, AND DEFECTS 4 23 12 31
edfas.org 1 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 DEPARTMENTS Author Guidelines Author guidelines and a sample article are available at edfas.org. Potential authors should consult the guidelines for useful information prior to manuscript preparation. A RESOURCE FOR TECHNICAL INFORMATION AND INDUSTRY DEVELOPMENTS AUGUST 2023 | VOLUME 25 | ISSUE 3 edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS 2 GUEST EDITORIAL Peter Friedrichs 38 SPECIAL ISTFA 2023 PREVIEW Frank Altmann 40 ISTFA EXHIBITORS LIST 41 2023 PHOTO CONTEST 42 2023 VIDEO CONTEST 43 EDUCATION NEWS Navid Asadi 44 IN MEMORIAM 46 DIRECTORY OF FA PROVIDERS Rosalinda Ring 47 TRAINING CALENDAR Rosalinda Ring 49 LITERATURE REVIEW Michael R. Bruce 51 PRODUCT NEWS Ted Kolasa 54 GUEST COLUMN Chuan Zhang 56 ADVERTISERS INDEX For the digital edition, log in to edfas.org, click on the “News/Magazines” tab, and select “EDFA Magazine.” Four-dimensional Scanning Transmission Electron Microscopy: Part I: Imaging, Strain Mapping, and Defect Detection Aaron C. Johnston-Peck and Andrew A. Herzing 4D-STEM provides researchers with information that can be analyzed in a multitude of ways to characterize a sample’s structure, including imaging, strain measurement, and defect analysis. 4 12 Failure Analysis of Photonic Integrated Circuits Frieder H. Baumann, Brian Popielarski, Ryan Sweeney, Felix Beaudoin, and Ken Giewont This article introduces silicon photonics, describes what is needed for photonics FA, and shows examples of FA in modern silicon photonics circuits. 23 Advancements in Image Pattern Recognition for LIT Hotspot Detection and Classification with Supervised Learning Kyu Kyu Thinn, Rui Zhen Tan, Teh Tict Eng, and Ming Xue The image pattern recognition algorithm for detecting LIT hotspots benefits image processing and can be leveraged to automate FA processes. 12 4 Multilayer Perceptron Development to Identify Plastics using Fluorescence Lifetime Imaging Microscopy Georgekutty Jose Maniyattu, Eldho Geegy, Maximilian Wohlschläger, Nina Leiter, Martin Versen, and Christian Laforsch Plastics can be classified by neural networks that are trained, validated, and tested by frequency domain fluorescence lifetime imaging microscopy measurements. 31 ABOUT THE COVER MOSFET leakage caused by stacking faults and aluminum spiking, both of which have punched through the junction. A break of the drain contact barrier layer allowed the aluminum to diffuse preferentially along the pair of intersecting stacking faults. This DF STEM image is reminiscent of Darth Vader in simmering rage. Photo by Wentao Qin, Onsemi, First Place Winner in False Color Images, 2022 EDFAS Photo Contest. 23 31
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 2 PURPOSE: To provide a technical condensation of information of interest to electronic device failure analysis technicians, engineers, and managers. Nicholas Antoniou Editor/PrimeNano nicholas@primenanoinc.com Mary Anne Fleming Director, Journals, Magazines & Digital Media Joanne Miller Senior Editor Victoria Burt Managing Editor Allison Freeman Production Supervisor ASSOCIATE EDITORS Navid Asadi University of Florida Guillaume Bascoul CNES France Felix Beaudoin GlobalFoundries Michael R. Bruce Consultant David L. Burgess Accelerated Analysis Jiann Min Chin Advanced Micro Devices Singapore Edward I. Cole, Jr. Sandia National Labs Michael DiBattista Varioscale Inc. Rosine Coq Germanicus Universitié de Caen Normandie Szu Huat Goh Qualcomm Ted Kolasa Northrop Grumman Space Systems Rosalinda M. Ring Thermo Fisher Scientific Tom Schamp Materials Analytical Services LLC David Su Yi-Xiang Investment Co. Martin Versen University of Applied Sciences Rosenheim, Germany FOUNDING EDITORS Edward I. Cole, Jr. Sandia National Labs Lawrence C. Wagner LWSN Consulting Inc. GRAPHIC DESIGN Jan Nejedlik, jan@designbyj.com PRESS RELEASE SUBMISSIONS magazines@asminternational.org Electronic Device Failure Analysis™ (ISSN 1537-0755) is published quarterly by ASM International®, 9639 Kinsman Road, Materials Park, OH 44073; tel: 800.336.5152; website: edfas. org. Copyright © 2023 by ASM International. Receive Electronic Device Failure Analysis as part of your EDFAS membership. Non-member subscription rate is $175 U.S. per year. Authorization to photocopy items for internal or personal use, or the internal or personal use of specific clients, is granted by ASM International for libraries and other users registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided that the base fee of $19 per article is paid directly to CCC, 222 Rosewood Drive, Danvers, MA 01923, USA. Electronic Device Failure Analysis is indexed or abstracted by Compendex, EBSCO, Gale, and ProQuest. The implementation of wide-bandgap-based power semiconductor solutions, in particular those based on SiC and GaN, has grown substantially over the last few years. Driving forces behind this market development include global megatrends like energy saving, decarbonization, and effective use of scarce resources. SiC power components offer distinct advantages. The new wide-bandgap technology is more than an evolutionary step forward, as we have seen in previous years with each new generation of silicon power devices; this technology has the capability to be a real game changer. SiC-based systems can be characterized by a steep change in performance, which can make them attractive for designers targeting innovative and disruptive solutions. To understand the differences between Si and wide-bandgap materials like SiC or GaN solutions, a closer look into the physical parameters of these new semiconductor materials, mainly the larger band gap and its implications, is required (see Fig. 1). Due to the much higher critical breakdown field of the material, the voltage range for fast and unipolar Schottky diodes as well as field effect based SiC switches (MOSFET, junction field effect transistor (JFET)) can be therefore extended to values well above 1000 V. Insulated gate bipolar transistors (IGBTs) or super-junction MOSFETs in combination with SiC diodes have already become the norm in various applications, such as solar, chargers, or power supplies. This combination, a fast silicon-based switch matched with a SiC diode, is often termed a hybrid solution. SiC transistors are on the way to becoming an alternative to today’s established IGBT technologies in industrial and automotive power electronics. A powerful SiC switch offering a proven and established ruggedness and AUGUST 2023 | VOLUME 25 | ISSUE 3 A RESOURCE FOR TECHNICAL INFORMATION AND INDUSTRY DEVELOPMENTS ELECTRONIC DEVICE FAILURE ANALYSIS GUEST EDITORIAL SiC POWER DEVICES AND RELATED ROBUSTNESS AND RELIABILITY ASPECTS Peter Friedrichs, Infineon Technologies AG peter.friedrichs@infineon.com edfas.org Friedrichs (continued on page 50) Fig. 1 Important physical parameters of modern power semiconductor materials.
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)
edfas.org 5 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 cluster (Fig. 2b), or a few fragmented signals due to cluttering by the circuitry layout or background contrast (Fig. 2c). In some cases, further challenges may include the presence of noise that may be associated to the setup environment, or the detection of very weak thermal signals. To manage data variability and noise removal, the algorithm implements the following four main steps: (1) detection of colored pixels from the gray background using Euclidean distance, (2) small particulate noise removal at pixel and connected component level, (3) clustering of nearby connected objects using density-based spatial clustering of applications with noise (DBSCAN),[4] and (4) identification of significant hotspots through ranking by size threshold based on the information from each image. The results of the algorithm based on 103 images containing 116 hotspots are discussed. For 86 images (83.5%), the hotspots were correctly identified as the only hotspot in the images. For 11 images (10.7%), the correct hotspots were identified along with other spurious hotspots. For six images (5.8%), the correct hotspots were not identified. In terms of hotspot detection, the algorithm achieves a sensitivity of 94.8% and has a false discovery rate of 17.9% (Fig. 3). Fig. 2 Three examples of images where the hotspot appears as (a) a single connected spot, (b) a diffused cluster, and (c) a fragmented spot that is broken up by the dark background. (b) (a) (c) Fig. 3 Results based on 103 images containing 116 hotspots. For the 17 images with incorrect hotspots detected, it was found that the spurious hotspots were big in size and thus, they were not removed during the noise removal and ranking by size steps. These hotspots were either comparable in size to the actual hotspots, hence, were identified along with the actual hotspots in 11 outcomes (Fig. 4a) or were larger than the actual hotspots, causing them to be selected in preference over the actual hotspots during ranking in the remaining six outcomes (Fig. 4b). It was noticed that the spurious hotspots were located outside of the package or die region. Therefore, the algorithm could be enhanced by implementing human-assisted identification on the region of interest before the application. IMAGE CLASSIFICATION WITH SUPERVISED LEARNING FA analysts typically reference the historical LIT images from the database to look for similar images. To improve the efficiency of image searching and automate image retrieval and ranking, a supervised algorithm has been developed to display similar images to a new query LIT image (Fig. 5). There are three main steps involved in this process: (1) algorithm for image similarity, (2) algorithm for image classification, and (3) re-ranking of images incorporating results from both image similarity and image classification algorithms. To work with a small dataset having only a few hundred images, a transfer learning method is used. Image features are extracted from the images by applying the pre-trained VGG16 network.[5,6] Principal component analysis (PCA) is further performed to construct image signatures that are relevant to this dataset. The top 100 significant PCA
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 6 components are taken out from the 4096 features out- put from VGG16. By implementing transfer learning together with PCA, the model is optimized for the small set of images. A two-layer classifier, designed to replicate the decision-making process of the FA analyst, has been developed to further improve the ranking algorithm. Upon receiving a query image, the first layer of the classifier differentiates between whether the image relates to the package or die level (Fig. 6). The second layer of the classifier then categorizes the image under the appropriate package class (if it is captured at the package level) or die class (if it is captured at the die level). This dataset contains five package classes (Fig. 7) and 16 die classes. Next, the algorithm extracts relevant images captured at the same package/die level belonging to the same package/die class. The classified images are then ranked according to the Euclidean distance from the query image. The Euclidean distances of images that belong to the same Fig. 4 Two examples where the hotspot recognition did not work. (a) The hotspot was correctly identified (near the center region) along with noise (at the edge). (b) The wrong hotspots (in red bounding boxes) were identified. The actual hotspot (in yellow circle) was not identified by the algorithm. (b) (a) Fig. 5 The methodology. The outline of each image corresponds to its class identity. Fig. 6 Layer 1 classifier determines whether images are taken at package-level (top) or die-level (bottom).
edfas.org 7 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 two classes as the query image are reduced by a fixed constant. Results from both layers of classifiers are used to rank higher to differentiate with those from a different class. Finally, the images are re-ranked and recommended to the FA analyst. The algorithm is tested on a dataset of 372 images that is divided into two subsets. 80% of dataset (298 images) are used to construct the database and the other 20% of dataset (74 images) are used as query images to evaluate the training and testing accuracy of the algorithm. Among (a) the 298 images, 176 images and 122 images belong to the package and die level, respectively. For the evaluation, the precision “k” is fixed at five, in other words, the top five recommendations from the number of relevant search results are selected. Table 1 shows the optimal number of PCA components and the corresponding cross-validation accuracies for the two-layer classifier. By selecting 10 PCA components as the first layer classifier, the highest accuracy attained is 92.2%. For Layer 2 package classifier, 20 PCA components are required to achieve an accuracy of 77.6%. For Layer 2 die classifier, 50 PCA components are required to achieve an accuracy of 47.5%. More PCA components are required to differentiate the finer features in each image. Higher accuracy is attained with larger number of training images and smaller number of classes. For instance, (b) (c) (d) (e) Fig. 7 Layer 2 package classifier differentiates the five package types (DIP, DSO, VQFN, LQFP, and BGA). Table 1 Cross-validation accuracy for Layer 2 classifier Classifier Optimal number of PCA components Cross-validation accuracy Layer 1 10 92.2% Layer 2 package 20 77.6% Layer 2 die 50 47.5%
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 8 Layer 1 classifier consists of 298 images for training two classes whereas Layer 2 classifiers consist of 175 images for training five package classes and 122 images for training 16 die classes. Table 2 shows the performance of the algorithm on the images in the database (training set) and the query images (testing set). The algorithm is reviewed with and without the classification step to understand whether the twolayer classifier is beneficial in providing relevant images. From the observation at Layer 1, precision above 90% is achieved for both the training and testing set without using classification. This implies that the image similarity algorithm alone can filter images at the same package or die level as the query image. Classification improves the precision slightly. At Layer 2 Package, precisions of 64.8% and 72.7% are obtained for the training and testing set respectively without using classification. At Layer 2 Die, precisions of 41.7% and 45.6% are obtained for the training and testing set respectively without using classification. For this Layer 2 Package and Die, the classification delivers significant improvement in the performances (e.g., the training set (package) has 16.1% improvement and the training set (die) has 24.3% improvement). PROGRESS ON SIGNAL-TO-NOISE RATIO Further progress can be made to improve the signal-tonoise ratio by reducing the noise at the raw LIT image with software noise reduction. Different methods to suppress noise to improve the quality of the images are studied. These include characterizing the amplitude and phase of raw images available from LIT and the effects of exposure time on the signal-to-noise ratio (Fig. 8). With the suppression for noise, hotspots detection using the proposed algorithm can be accomplished at a shorter exposure time. CONCLUSION The image pattern recognition algorithm for detecting LIT hotspots not only benefits image processing, it can be leveraged to automate FA processes that require the identification of anomalies similar to hotspots in images as demonstrated in this work. Image classification with supervised learning has also shown favorable results. The image similarity search extracts image features and ranks them, while the image classification refines the ranking based on relevance for FA analysts. The precision of the algorithm is significantly improved after the classification process. To improve precision further, a larger dataset can be used. Fig. 8 (a) Original raw LIT image (b) after enhancement on image (a). (b) (a) Table 2 Precision for training and testing set with and without classification Precision, % Layer 1 Layer 2 Package Die Training precision without classification 92.0 64.8 41.7 Testing precision without classification 97.0 72.7 45.6 Training precision with classification 95.1 80.9 66.0 Testing precision with classification 97.3 76.2 54.3
edfas.org 9 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 There has been promising progress in improving the signal-to-noise ratio by processing the amplitude and phase raw images and acquiring raw image data at different exposure times. In future studies, the aim is to develop algorithms that are sensitive enough to detect hotspots at varying exposure times. ACKNOWLEDGMENTS This research work is funded under SIT-MOE Ignition Grant, R-MOE-A403-F022, Singapore Institute of Technology (SIT), Singapore. We would like to thank all SIT members (University Co-Principal Investigator: Associate Professor Neelakantam Venkatarayalu, Associate Professor Indriyati Atmosukarto and Associate Professor Benjamin Premkumar Annamalai) for their contributions. This collaboration between SIT and Infineon offers opportunity to students to embark on applied research projects on problem statements and digitalization arising from the industry. REFERENCES 1. L.C. Wagner: “Failure Analysis,” Handbook of Semiconductor Manu- facturing Technology, Second Edition, 2007, https://doi.org/ 10.1201/9781420017663. 2. R.Z. Tan, et al.: “Localization of Hotspots from Lock-in Thermography Images for Failure Analysis,” 2021 IEEE 23rd Electronics Packaging Technology Conference (EPTC), p. 45-49, 2021, https://doi.org/10.1109/ EPTC53413.2021.9663910. 3. R.Z. Tan, et al.: “Supervised Image Retrieval and Ranking Technique for Lock-in Thermography Images,” 2022 IEEE International Symposium on the Physical and Failure Analysis of Integrated Circuits (IPFA), 2022, https://doi.org/10.1109/IPFA55383.2022.9915757. 4. R.J.G.B. Campello, et al.: “Density-based Clustering,” WIREs. Data Min. Knowl. Discov., 2020, https://doi.org/10.1002/widm.1343. 5. K. Simonyan and A. Zisserman: “Very Deep Convolutional Networks for Large-scale Image Recognition,” 2014, https://doi.org/10.48550/ arXiv.1409.1556. 6. A. Krizhevsky, I. Sutskever, and G.E. Hinton: “ImageNet Classification with Deep Convolutional Neural Networks,” Advances in Neural Information Processing Systems 2012, 25. ABOUT THE AUTHORS Kyu Kyu Thinn works for Infineon Technologies Asia Pacific Pte. Ltd. as a failure analysis engineer with the Product Analysis Team in Singapore and received her bachelor’s in electrical engineering from National University of Singapore. She has more than 10 years of experience in analyzing electrical failures of IC devices. Rui Zhen Tan is an assistant professor in the engineering cluster at Singapore Institute of Technology, SIT. Her research interests are in data analytics, predictive maintenance, and mathematical modeling. Prior to SIT, she was a postdoc at the Bioinformatics Institute, A*STAR. Teh Tict Eng has a master’s degree in electrical and electronics engineering from National University of Singapore. She has many years of experience in semiconductor industry, focusing on failure analysis and people management. Her expertise is in quality management system in failure analysis lab scope. Ming Xue has 40 years of electronics/semiconductor industrial experience: 10 years as RF designer, four years as a print circuit board assembly process engineer, and 26 years as head of Singapore FA and senior principal. He is key technical staff in Infineon Failure Analysis and Back End. Advertise in Electronic Device Failure Analysis magazine! For information about advertising in Electronic Device Failure Analysis: Kelly Johanns, Business Development Manager 440.671.3851, kelly.johanns@asminternational.org Current rate card may be viewed online at asminternational.org/mediakit.
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 10
edfas.org 11 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 12 FOUR-DIMENSIONAL SCANNING TRANSMISSION ELECTRON MICROSCOPY: PART I: IMAGING, STRAIN MAPPING, AND DEFECT DETECTION Aaron C. Johnston-Peck and Andrew A. Herzing National Institute of Standards and Technology, Gaithersburg, Maryland aaron.johnston-peck@nist.gov EDFAAO (2023) 3:12-22 1537-0755/$19.00 ©ASM International® INTRODUCTION Scanning transmission electron microscopy (STEM) has become an indispensable tool for materials characterization due to its ability to elucidate structure and composition with nanometer scale or better spatial resolution. In the scanning transmission electron microscope, electromagnetic lenses form an electron probe that is scanned across a sample (Fig. 1a). As the electron probe interacts with the sample, a distribution of forward scattered electrons (or diffraction pattern) results (Fig. 1b). Within this diffraction pattern, several features can be found, all of which originate from different scattering mechanisms. For crystalline samples, the features that dominate these patterns include sharp, relatively intense discs or spots arising from zero order Laue zone diffraction or Bragg diffraction; Kikuchi bands arising from multiple scattering; or a broad, slowly changing background that is due to several mechanisms including thermal diffuse scattering. Because all these features carry information about the local nature of the sample being probed by the electron beam, electron diffraction patterns are a rich source of information for high spatial resolution materials characterization. However, many traditional imaging modes of STEM and TEM isolate a subregion in the diffraction plane and form an image using only those particular electrons. For example, electrons detected at high scattering angles produce contrast that displays an atomic number dependence, while forming an image using a single diffraction spot produces contrast sensitive to variations in lattice parameter (i.e., strain). In these examples, the images are formed by utilizing a limited fraction of the information generated from the specimen-electron interaction. A paradigm shift would be to detect and record all the scattering information generated and subsequently analyze that dataset to build images that reveal both composition and strain. A new suite of analytical techniques known collectively as fourdimensional scanning transmission electron microscopy (4D-STEM) is that paradigm shift. Keep in mind, 4D-STEM offers much more than the ability to generate “traditional” images. It provides the framework to produce two dimensional (as well as three dimensional) representations of strain, electromagnetic fields, crystallographic phase, and many other material characteristics. This article will introduce the basics of the technique and some areas of application with an emphasis on semiconductor materials. 4D-STEM is a spatially resolved electron diffraction technique that records the electron scattering distribution at each sampling point. The diffraction pattern is projected onto a two-dimensional pixelated recording device providing intensity information as a function of angle (kx, ky). Electromagnetic deflectors scan the electron probe over the sample in a two-dimensional grid of positions (rx, ry), while the two-dimensional recording device is synchronized with the scan to record the diffraction pattern at each position resulting in a four-dimensional dataset (rx, ry, kx, ky) (Fig. 1c). Taking a step back, microbeam electron diffraction, nanobeam electron diffraction (NBED), and convergent beam electron diffraction (CBED) have been used to describe methods where electron diffraction patterns are generated by illuminating a small sample area, but not necessarily coupled with spatially resolved sampling. While reports of spatially resolved diffraction experiments can be found in the literature from as early
edfas.org 13 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 the ability to map magnetic domains,[6] serving to highlight the promise of the technique rather than ushering in widespread adoption. The implementation of 4D-STEM relies on the availability of detectors with low noise and high readout speeds, single electron sensitivity in addition to adequate radiation hardness. Early attempts at developing or repurposing detectors with these characteristics were met with varying degrees of success.[7-12] However, it was not until the mid2010s when detectors targeting 4D-STEM applications became more common and the number of commercial products in the market has increased steadily since then. A more in-depth discussion on detector technology and development is beyond the focus of this article but can be found elsewhere.[13] The remainder of this article will introduce various 4D-STEM methods related to imaging, Fig. 1 A schematic of scanning transmission electron microscopy including both a single channel and a pixelated detector, where α is the convergence semi-angle, 2θB is the Bragg angle of the diffracted beam labeled hkl (a). A colorized electron diffraction pattern where different features can be observed, including Bragg reflections, Kikuchi bands, and a diffuse background (b). An example 4D-STEM dataset with 578 individual diffraction patterns from GaN [12 _ 10]. The region of GaN, indicated by the ball model where Ga is green and N is yellow, was sampled in a 32 by 18 grid where each cross indicates a point in the scan (c). The diffraction patterns were simulated using the multislice method as implemented in the software program Dr. Probe.[56] (a) as the mid-1980s,[1] where two-dimensional diffraction patterns were recorded along a one-dimensional scan. 4D-STEM experiments were not far behind, with possibly the first demonstration being reported in 1989.[2] Initial experimentation was motivated by attempts to reconstruct real-space images from diffraction data and potentially surpass the image resolution of the microscope imposed by the lens aberrations.[3,4] These efforts culminated in a reconstructed image with a point resolution that far exceeded the information limit of the microscope,[5] and while incredibly impressive, these experiments required herculean efforts. The stability of the microscopes, detector technologies, and computing resources available at that time were inadequate to make this type of measurement practical. As a result, early 4D-STEM measurements amounted to isolated experiments, such as demonstrating (b) (c)
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 14 strain mapping, and defect detection. Because the field of 4D-STEM is continually evolving, we include some directions which are still in development. A follow-up article including the topics of mapping crystallographic orientation, phase, and electromagnetic fields is in preparation. IMAGING STEM images—be they bright field (BF), annular bright field (ABF), or high-angle annular dark field (HAADF)—are traditionally generated by single channel integrating detectors with either an annular or circular geometry. The detector in the far-field subtends a solid angle established by the physical dimensions of the detector and the post-specimen optics of the microscope. Electrons scattered into that solid angle generate a signal, which is subsequently processed and digitized. Then a numerical grayscale value is assigned to an image pixel corresponding to each position of the scan. Alternatively, this process can be mimicked with a 4D-STEM dataset and summation. For each diffraction pattern within the 4D dataset, a subset of pixels can be summed, likewise producing a numerical value at each point in the raster, producing what is sometimes referred to as a virtual[14] or synthetic image.[15] This process has been successfully demonstrated for BF and ADF images,[10] as well as reconstructing images with atomic resolution.[16] Quantitative agreement between images reconstructed from 4D-STEM datasets with simulations was found[17] indicating that virtual images can replicate the fidelity of images acquired using traditional detectors. The flexibility of generating images from a 4D dataset is what differentiates it from a traditional single channel STEM detector. On a microscope, there are typically one to three STEM imaging detectors. Their geometries are rudimentary (albeit practical and chosen with good theoretical underpinning), the number of collection solid angles available is limited by the combination of detector dimensions and discrete predefined camera lengths (although this can be extended through various modifications of the post-specimen optics, e.g., inserting a post-specimen aperture to limit the outer collection angle[18]), and only one detector can collect a particular angle. By comparison, when using the 4D-STEM dataset, a limitless set of virtual detectors can be applied and the collection angles do not need to be unique due to limitations of detectors subtending one another. In Fig. 2, a set of four reconstructed images are shown from GaN [12 _ 10], including a complex geometry and the reuse of the same collection angles. Moreover, it is convenient to apply virtual detectors that highlight the presence of specific features. For instance, superlattice and Bragg reflections associated with a particular phase can be selected to generate position maps of minority phases within a matrix.[14,19] This may sound similar to the process of forming a dark field TEM image. Indeed, the similarity can be understood through the principal of reciprocity in TEM and STEM which can be crudely summarized as follows: In a microscope, if a signal is detected at a point A and a source is placed at point B, then the same signal would be detected at B if the source were placed at A.[20,21] The collection angle defined by the objective aperture in TEM and the convergence angle defined by the probe forming aperture in STEM are analogs in this situation. Yet, the flexibility of convergence lens systems allows for a wide range of convergence semi-angles including to very small (< 1 mrad) values, an angular range that may not be accessible with the objective apertures that are commonly installed in TEMs, which ensures the separation of Bragg Fig. 2 A demonstration of atomic resolution virtual imaging of GaN orientated to the [12 _ 10] zone axis. An individual diffraction pattern from the 4D-STEM dataset and corresponding ball model of GaN (a). Reconstructed images with the corresponding mask are shown (b). The color bar is fraction of the incident beam intensity (unitless). All images were generated from the same dataset. (a) (b)
edfas.org 15 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 diffraction reflections even in structures with large unit cells or multiphase materials. Image generation is not limited to integrating a subset of pixels within the 4D dataset. Matched illumination detector interferometry STEM (MIDI-STEM) uses a patterned phase plate with concentric rings to form the electron probe in lieu of the traditional circular probe-forming aperture. This concentric ring pattern is transferred to the transmitted and diffracted beams in the resulting diffraction patterns. The MIDI-STEM image signal is then computed by taking the difference between the sum of all odd ring intensities and the sum of all even ring intensities. The result is an imaging mode that produces nominally linear phase-contrast, thereby enhancing the image contrast of weakly scattering materials, which is particularly useful if they are prone to electron beam-induced dam- age.[22] Similar approaches of recording a 4D dataset while utilizing a phase plate to modify contrast has been repor- ted.[23,24] The combined parameter space of 4D datasets and structuring the incident probe characteristics creates a large parameter space to enable information rich measurement schemes that would not be accessible otherwise. STRAIN The strain state of a material or the interface between two materials can have a large effect on the resulting properties. For example, the charge carrier mobility in silicon-based MOSFET devices is often finely tuned by carefully controlling the amount of lattice strain in the material which in turn alters the electronic band structure.[25] Defects can also arise when the strain state of a material becomes large, as the crystal will relax through the formation of stacking faults or dislocations. Therefore, it is critical for process control, failure analysis, and materials development to be able to quantitatively measure strain at high spatial resolutions, and fortunately many TEM-based methods exist for doing so. For crystalline materials, the lattice periodicity can be directly visualized using highresolution TEM (HRTEM) and relative deviations in lattice parameter can be extracted via comparison of Fourier transforms calculated from various subregions in an image. This image-based approach is simple, but quantification can be problematic. Diffraction methods are much preferred for high-precision measurements and there are two primary approaches that are commonly employed: CBED and NBED. The CBED method relies on analyzing the features due to higher-order Laue zone diffraction in images of a highly-convergent electron probe.[26,27] Strain measurement via CBED has the advantage of high precision (values as high as 2 x 10-4 have been reported[28]) and can provide an absolute measure of the lattice parameter present in a material. However, to achieve this precision the specimen must be made very thin so that multiple scattering can effectively be ignored or, alternatively, an energy filter must be employed. Further, the analysis of the diffraction features can be complex and difficult to automate, which makes performing the measurement in a spatially resolved manner problematic. NBED, by contrast, is based on measuring the displacement of the diffraction spot and is well-suited to spatially resolved strain measurements with a slight tradeoff in precision. The resulting 4D dataset is then processed to determine the diffraction disc spacings and inversely the lattice spacing, at each probe position. NBED is conceptually simple and in theory should easily provide a straightforward, high-precision method for measuring strain at subnanometer-length scales. However, there are various challenges with data processing that must be carefully addressed to achieve the best result. The primary difficulty is in the measurement of the diffraction disc locations. In contrast to conventional electron diffraction where the diffraction maxima are strongly peaked and easily located, the use of a convergent probe in NBED distributes the diffracted intensity over a disc in reciprocal space whose diameter is defined by the angular extent of the incident probe. The best-case scenario is for very thin specimens where the contrast variation within the transmitted and diffracted discs will be minimal due to the lack of dynamical effects. In these cases, determination of the diffraction disc centers can be done via simple circle detection algorithms or by cross correlation of a template image of the transmitted central disc through vacuum. As the specimen thickness increases, these approaches to disc position detection become problematic as the intensity in each disc becomes less uniform and varies non-systematically. To overcome these problems, a number of different algorithms, including machine learning, have been employed to improve precision of disc position identification.[29-34] Several other approaches have been pursued to make NBED-based strain measurements more robust, including the use of patterned probe forming apertures.[35,36] Aperture shapes such as crosses or bullseyes are overlaid on the transmitted and diffracted discs in the NBED patterns, thus making centroid determination less ambiguous and sensitive to contrast variation in the discs. Significant improvements in precision of the measurement have been reported. However, it has also been shown by Mahr et al.[37] that the patterned apertures produce distorted probes with extended tails which can sample parts of the
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 16 specimen that are far from the central probe position. Near interfaces this results in a decrease in accuracy of the strain measurement as the resulting diffraction patterns will contain contributions from both sides of the boundary. Another method for improving the precision of NBED disc position determination is precession electron diffraction (PED). In PED, the incident electron probe is tilted slightly off the optic axis and then azimuthally rotated (precessed) from 0° to 360° forming a cone of illumination. The diffraction pattern collection time is set to be equal to an integer multiple of the precession time. The effect of the precession process is to significantly dampen the effects of dynamic scattering. In turn, this results in a reduction of intensity variation in the diffraction discs and makes the determination of their location much more precise. While this method was not commonly available on most commercial microscopes in the past, this technique has seen widespread adoption in recent years and can be seamlessly integrated into the microscope acquisition interface. NBED combined with PED has been used to produce strain maps of Si/SiGe multilayer structures as well as from real semiconductor devices with very high precision (Fig. 3 panel i).[38] In the latter case, the strain field in a 22 nm transistor channel near recessed SiGe source and drains was mapped with precision on the order of 2 x 10-4. An alternative method for treatment of NBED data, which also seeks to overcome difficulties with disc position determination, has recently been developed and was inspired by cepstral analysis in audio signal processing. The so-called exit-wave power cepstral (EWPC) transformation[10] converts each reciprocal space diffraction pattern in an NBED dataset to a real-space pattern via Fourier transformation. This results in patterns with sharp peaks of uniform intensity as compared to the discs with heterogeneous contrast in the NBED patterns. The location Fig. 3 Examples of strain mapping. Panel i reproduced from Rouviere,[38] shows a strain analysis of a transistor device. (a) Virtual BF image of the transistor region showing the 22 nm channel (C) and the recessed SiGe source (S) and drain (D). (b)-(e) Strain maps calculated from N-PED patterns: εyy, εxx, εxy, and rotation (left to right). Panel ii reproduced from Ozdol,[40] NBED analysis of GaAs/GaAsP multilayer device cross-section. (a) HAADF image of the analyzed section. (b) Representative nanodiffraction pattern taken from the dataset with the [200] and [220] diffraction vectors that were analyzed for the strain analysis highlighted. (c) NBED strain map of εyy strain where y is the [002] direction. (d) NBED strain map of εxx strain where x is the [220] direction.
edfas.org 17 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 of the EWPC peaks can be determined far less ambiguously and therefore the precision of the strain measurement is improved. This method has been used to map the strain state in core-shell nanoparticles and to visualize domain distributions in ferroelectric materials.[39] While NBED has been used for many years for spatially resolved strain measurements, the development of high-speed direct detection cameras and 4D-STEM has increased the capabilities and ease of data collection for this technique. This has two major advantages. First, the high speed of acquisition allows for strain analysis over much larger fields of view with no loss in spatial resolution. Second, the high sensitivity of these detectors allows data of similar quality to be collected using a far lower electron dose. The advantage of an increased field of view for strain mapping has been demonstrated by Ozdol et al. in analyzing a GaAs/GaAsP multilayer specimen (Fig. 3 panel ii).[40] While maintaining a spatial resolution of 1 nm, they were able to map the in-plane strain fields over a 1 µm field of view with a precision of 1 x 10-3. The high speed of 4D-STEM-based strain analysis can also be leveraged to carry out in situ experiments where the strain distribution in a specimen can be monitored as a function of external stimuli such as mechanical deformation or thermal cycling. Pekin et al.[41] used this approach to study strain development in stainless steel using a holder capable of pulling the specimen in tension during STEM analysis. With this setup, they were able to simultaneously collect real-space images and reciprocalspace diffraction patterns to monitor dislocation creation and lattice expansion directly. Further, the introduction of pump-probe experimental configurations can extend this technique into the realm of sub-ns dynamic processes. For example, the time and spatially dependent strain state of single crystalline silicon patterned with tungsten discs was measured as acoustic waves propagated from the tungsten upon absorbing energy from laser pulses.[42] Finally, since STEM is inherently a projection-based technique, all the applications mentioned thus far in this section involve two-dimensional mapping of various aspects of the strain field perpendicular to the incident beam (i.e., in the plane of a thin specimen) and the value measured at each point is an average over the thickness of the sample. Typically, four components of the deformation tensor can be derived: these are the strain in the x- and y-directions as well as the in-plane shear and rotational components. Retrieving additional components of the tensor requires either tilting the specimen to make measurements at different orientations or analysis of higher order Laue zone (HOLZ) information. The high-speed and sensitivity of 4D-STEM may enable the use of electron tomographic techniques to measure the full, three-dimensional strain tensor at high spatial resolution at each threedimensional position within the specimen.[43] In electron tomography, multiple images are collected as the sample is tilted with respect to the electron beam. This image series is then used to calculate a three-dimensional image of the material. By collecting a full 4D-STEM NBED dataset at each specimen orientation, the projected strain maps can potentially be used as input for a three-dimensional reconstruction algorithm to recover the information lost in projection. This reconstruction process is non-trivial and development is on-going. DEFECTS Defects have significant implications for device performance and having the tools to reliably detect, categorize, and characterize them is crucial. Because the structure and characteristics of defects are incredibly diverse, spanning a range of length scales from macroscopic bulk defects to zero-dimensional point defects, a variety of characterization techniques are needed. Two such cases include images acquired under two-beam conditions to identify the Burgers vector of dislocations and HAADFSTEM images to reveal the position of individual, high atomic number dopant atoms in a low atomic number matrix.[44] As we previously discussed, 4D-STEM can reproduce traditional methods of defect characterization through virtual imaging, but more importantly, it can also be used to enhance existing methods or even enable new techniques. Case in point, Shao et al. have shown that a cepstral analysis of 4D-STEM data can be used to separate electron diffuse scattering from Bragg diffraction. This is performed by calculating the difference between the cepstral transforms of each individual diffraction pattern and the region averaged diffraction pattern.[45] The result is a spatial map of the diffuse scattering produced at each point in the sample that results from thermal or static atomic displacements of the periodic crystal lattice.[20] This technique, while recently developed, has already been used to identify the distorted region around a dislocation core in a SiGe thin film,[45] chemical short range order in a medium-entropy CrCoNi alloy,[46] and phase domains in MnO2 undergoing an intercalation reaction. [47] Some methods for defect characterization are adaptations of existing electron microscopy techniques that benefit from the spatially resolved sampling of 4D-STEM. Structure factor refinement is one such approach. A structure factor describes the scattering of an electron (continued on page 20)
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 18
edfas.org 19 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3
edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 20 wave by the unit cell of the crystal structure, meaning it is sensitive to the types of atoms present and their arrangements within the unit cell. Refinement, i.e., a comparison of simulated and experimental data to quantify the structure factor that best reproduces the experimental results, of energy filtered diffraction patterns (energy filtering removes inelastic scattering which improves the fidelity of the refinement) was used to quantify the concentration of nitrogen in dilute-nitride GaNxAs1-x epilayers as a function of position.[48] It also provided insight into whether N incorporated interstitially or substitutionally, as well as in identifying the presence of As vacancies. Quantitative refinement of structure factors can also be used to understand bonding through the generation of charge density maps.[49,50] Yet to date, most structure factor refinement using CBED has been carried out on bulk crystals intentionally positioning the electron probe away from defects, however, a spatially resolved method to understand changes to bonding in proximity to defects has been proposed.[51] Other new 4D-STEM techniques can provide high spatial resolution complements to measurements without site specificity. For example, dilatometry or positron annihilation spectroscopy have traditionally been used to measure bulk averages of vacancy concentrations, it has recently been shown that 4D-STEM can be used to monitor more local changes in point defect concentrations.[52] This was done by directly correlating independent measures of changes in lattice parameter and the volumetric expansion as external stimuli are applied (Fig. 4 panel i). Changes in defect concentrations were monitored in regions of approximately 100 nm2 during in situ heating of gold or as a function of electron radiation in aluminum films.[52] Because defects perturb the long-range order of a crystal and therefore the electron scattering process, several approaches look to analyze the symmetry of diffraction patterns to identify defects. One such method is coined as symmetry-STEM (S-STEM) and is based on the cross correlation of diffraction patterns with themselves after a symmetry operation (e.g., rotation or mirror) has been applied. In the analysis, intensity maxima occur when the symmetry operation is satisfied and intensity decreases at symmetry-breaking events. It was demonstrated that this approach can track changes in symmetry with atomic resolution (Fig. 4 panel ii). Because the image contrast of this correlation map can be interpreted as deviations from a specific symmetry operation, this approach can be used to identify regions of the crystal Fig. 4 Panel i reproduced from Mills.[52] Vacancy concentration mapping of Au as it is heated, independent measurements of lattice parameter and length quantified the increase in vacancies (ΔN/N). Panel ii reproduced from Krajnaka.[53] An example of symmetry-STEM analysis on [100] CeB6. S-STEM images for different applied-symmetry operations are compared with the traditional imaging modes of BF-, ABF-, and HAADF-STEM. FOUR-DIMENSIONAL SCANNING TRANSMISSION ELECTRON MICROSCOPY (continued from page 17)
www.asminternational.orgRkJQdWJsaXNoZXIy MTMyMzg5NA==