Scanning acoustic microscopy (SAM) is a critical nondestructive technique for detecting buried defects in semiconductor devices [1]. However, interpreting the complex acoustic data requires expert knowledge, especially as microelectronic structures grow increasingly intricate. This study presents a novel, unsupervised approach for defect detection in SAM data using an autoencoder neural network trained exclusively on intact reference samples. The method exploits reconstruction error as an indicator of signal anomalies, identifying potential defects without prior knowledge of their characteristics or the need for defective training samples. The resulting similarity parameter facilitates intuitive visualization of outlier regions through overlay on high-resolution SAM images. Furthermore, by analyzing the spatial and temporal deviation of the reconstructed signals, the method enables localization of the signal interval from which the deviation originates. This provides a rough estimation of the defect depth, adding an additional layer of interpretability to the detected anomalies. This enhances defect detection sensitivity and operator support, making SAM more accessible for less experienced users and highly promising for automation in production environments. While the method effectively highlights signal deviations, it may also flag non-defect-related anomalies such as sample tilting. Nonetheless, it significantly advances automated and operator-independent semiconductor failure analysis by enabling reliable detection of structural anomalies through learned representations of intact signal patterns.

Die Analyse und Prognose von Hochwasserereignissen erfordert kalibrierte Modelle, welche auf umfangreicher und hochgenauer Datengrundlage basieren. Unbemannte Luftfahrzeuge (UAV) in Kombination mit Künstlicher Intelligenz haben die Möglichkeiten in der flexiblen, hochfrequenten Datenerfassung signifikant erweitert. Mit SSegRef2Surf wurde ein Werkzeug entwickelt, welches automatisiert Wasserflächen aus Orthophotos klassifiziert, durch den Verschnitt mit Geländedaten verfeinert und geometrische Attribute (Wasserspiegellage und Wassertiefe) ergänzt.

Scanning Acoustic Microscopy (SAM) is a critical nondestructive technique for detecting buried defects in semiconductor devices [1]. However, interpreting the complex acoustic data requires expert knowledge, especially as microelectronic structures grow increasingly intricate. This study presents a novel, unsupervised approach for defect detection in SAM data using an autoencoder neural network trained exclusively on intact reference samples. The method exploits reconstruction error as an indicator of signal anomalies, identifying potential defects without prior knowledge of their characteristics or the need for defective training samples. The resulting similarity parameter facilitates intuitive visualization of outlier regions through overlay on high-resolution SAM images. This enhances defect detection sensitivity and operator support, making SAM more accessible for less experienced users and highly promising for automation in production environments. While the method effectively highlights signal deviations, it may also flag non-defect-related anomalies such as sample tilting. Nonetheless, it significantly advances automated and operator-independent semiconductor failure analysis by enabling reliable detection of structural anomalies through learned representations of intact signal patterns.

The threat posed by counterfeit electronics has grown since the 2020 supply-chain shutdown, increasing the necessity for a fast, non-destructive method that flags suspect parts with high confidence. This paper describes a custom, low-cost, two-dimensional near-field-scanning system that uses self-built electromagnetic probes and streamlined signal-processing software to identify relabelled or cloned integrated circuits by their unique emission footprints. Case-study measurements on known relabelled and cloned samples of type STM32 and FTDI with verification by X-Ray imaging demonstrate the setup’s effectiveness and simplicity for routine counterfeit screening.

Through the mobilization of movable objects due to the extreme hydraulic conditions during a flood event, blockages, damage to infrastructure, and endangerment of human lives can occur. To identify potential hazards from aerial imagery and take appropriate precautions, a change detection tool (CDT) was developed and tested using a study area along the Aisch River in Germany. The focus of the CDT development was on near real-time analysis of point cloud data generated by structure from motion from aerial images of temporally separated surveys, enabling rapid and targeted implementation of measures. The differences identified in the study area using distance comparison (M3C2) were segmented into individual components and categorized. Subsequently, the data was compared to existing two-dimensional hydrodynamic numerical calculation results (HQ100). The implementation of the CDT is feasible for a variety of RGB camera-equipped aerial vehicles due to the point cloud-based analysis and postprocessing. By overlaying and visualizing the detected changes with numerical simulation results, a quick assessment of the hazard potential in the event of a possible flood can be made. In the case of the study area along the Aisch River, the localization of construction materials, a steel container with debris pile, and a motor vehicle in the flood hazard zone of a potential HQ100 event could be confirmed, although no mobilization of the materials was to be expected due to the expected hydraulic conditions of a flood event.

Effective response to flood events requires high-resolution, frequently updated data on flooded areas for comprehensive flood risk assessments. Unmanned aerial vehicles (UAVs) equipped with conventional camera systems and classification based on orthophotos from photogrammetric postprocessing and artificial intelligence are widely used to detect flooded areas. However, these methods often involve time-intensive pre- and postprocessing steps and fail to incorporate geometric factors such as elevation data and water depths. This study introduces SSegRef2Surf, a novel tool that integrates classified flood raster data with terrain information. SSegRef2Surf refines and optimizes coarse raster classifications by filling shadowed areas and correcting misclassified regions. This tool reduces data requirements for AI training and minimizes postprocessing time, enabling near real-time flood monitoring. All processes necessary for SSegRef2Surf were optimized through sensitivity and accuracy analyses to reduce postprocessing duration to a minimum. A comparison of the SSegRef2Surf results with two-dimensional (2D) numerical model results for a flood event revealed discrepancies in the 2D model, caused by inaccuracies in the underlying terrain data. This comparison showed that 30% of the flooded areas identified in the 2D numerical results were incorrect, while missing areas (11%) were added. This highlights the significant potential of SSegRef2Surf for near real-time flood monitoring and traceability of flood events, as combining UAVs’ high-frequency surveying capabilities with SSegRef2Surf allows for more effective validation and optimization of 2D models.

Presentation slides for the ISTFA 2025 Tutorial session “Machine Learning in Semiconductor Failure Analysis: Techniques and Case Studies.”

Various complementary measurement systems are available for determining sediment layers in dam reservoirs. This study evaluated and compared a selection of these systems across reservoirs of varying scales. In addition to establish applica bility limits, the delimitation of the necessary data quantity in the field of tension between processing feasibility and suf ficient data precision were a core goal of the research. Due to the results of the studies a conceptual framework for survey ing of sediments in dam reservoirs was developed. The scope of the concept was to unify the data acquisition and processing in terms of different databases and investigation objectives.

In development and production of microelectronics products the assessment of the condition of either the full component or only specific parts is of high relevance. To allow for screening and monitoring and to leave the part unaltered the inspection techniques are required to operate non-destructively. While this allows for full-cover final inspection it also enables repetitive monitoring beneficial for the exploration of material interactions, potential subsequent defect formations and consequently for failure isolation. With increasing complexity inherent in advancing microelectronic technologies, high reliability, robustness and thus, successful failure analysis is essential. The correspondingly shrinking feature sizes and the involvement of heterogeneous materials highly challenge existing techniques. Furthermore, the interpretation of the acquired data becomes increasingly difficult requiring specifically skilled operating personal. In previous studies machine learning (ML) approaches have been developed and evaluated for their ability to analyze signals acquired by scanning acoustic microscopy (SAM) with the goal of automated defect detection, characterization, and failure isolation. The present paper investigates different ML architectures to analyze the time signals after transformation into the spectral- and wavelet domains. Results showed that 2D CNNs analyzing the acquired acoustic signals in the wavelet domain representation performed best, however at the expense of additional computational effort. Furthermore, ML-based analysis was explored for its potential to locate and isolate electrically active defects in the depth-dimension based on thermal emissions using lock-in thermography (LIT). Obtained LIT-related results are promising, however require further research to fully enfold its potential. It was further found that transfer properties of the inspection tools interfere with the defect specific signal features and thus so far tie the trained models to the specific equipment used. Future work should therefore focus on removing the specific tool related transfer characteristics of the equipment from the measurement data to allow for intra-tool compatibility and thus a more generalized application.

Localizing security-relevant hard blocks on modern System-on-Chips for physical attacks, such as side-channel analysis and fault attacks, has become increasingly time-consuming due to ever-increasing chip-area and -complexity. While this development increases the effort and reverse engineering cost, it is not sufficient to withstand resolute attackers. This paper explores the application of camera-based lock-in thermography, a nondestructive testing method, for identifying and localizing security hard blocks on integrated circuits. We use a synchronous signal to periodically activate security-related functions in the firmware, which causes periodic temperature changes in the activated die areas that we detect and localize via an infrared camera. Using this method, we demonstrate the precise detection and localization of security-related hard blocks at the die level on a modern SoC.

Basierend auf 3-D-HN-Untersuchungen des Ist-Zustandes am Hochwasserrückhaltebecken Watzdorf konnte festgestellt werden, dass im Falle extremer Hochwasserereignisse der entsprechende Zufluss voraussichtlich nicht schadlos abgeführt werden kann. Aufgrund dieser Erkenntnis wurde durch konstruktive Anpassungen im Bereich der Hochwasserentlastungsanlage ein Plan-Zustand entwickelt und numerisch untersucht, durch welchen eine Erhöhung des schadlos abführbaren Zuflusses erfolgen sollte.

Recent flood events (FE) in Germany have shown that the extent and impact of extreme flood events cannot be estimated solely based on numerical models. For analyzing the development of such an event and to develop and implement safety measures more efficiently, additional data must be collected during the event. Within the scope of this research, the possibilities of near real-time recording using an unmanned aerial vehicle (UAV) and data processing with the Structure from Motion (SfM) method were tested in a case study. Different recording parameter combinations were tested in the Laufer Muehle area on the Aisch river in Germany. The focus of the investigations was the identification of a parameter combination that allows a short recording interval for aerial imagery. Based on these findings, the identification of changes in the study area by comparing multitemporal photography (flood prevention), as well as the recording of flooded areas during a FE should be possible. The accuracy analysis of the different parameter combinations between two point clouds as well as the process of change detection was done by a Multiscale Model to Model Cloud Comparison (M3C2) and including ground control points. As a result, a parameter combination was identified which led to the desired results in the study area. The processes were transformed into fully automated and scripted workflows. The results serve as a basis for establishing a workflow for near real-time analyses in future studies.

Für die Erfassung von Sedimenthorizonten in Talsperren existieren komplementäre Messsysteme. Diese wurden an Talsperren angewendet und verglichen. Dabei spielte, neben der Festlegung von Grenzen für die Anwendbarkeit, vor allem die Abgrenzung des notwendigen Datenumfanges im Spannungsfeld von Verarbeitbarkeit und ausreichender Datenpräzision eine zentrale Rolle. In Erweiterung dessen wurde ein Messkonzept entwickelt, um die Aufnahme und Verarbeitung der Daten, trotz unterschiedlicher Datengrundlagen und Untersuchungsziele, zu vereinheitlichen.

3D heterogeneous integration (HI) and advanced packaging (AP) technologies require small volume, high-density interconnects for stacking discrete chips for which the reliability of interconnects becomes crucial. Intermetallic compounds (IMCs) based mu bumps have been shown to outperform solder-based mu bumps concerning their resistance to electromigration (EM) related failures, which is a key index to assess the interconnect reliability. Cu-Sn solid-liquid interdiffusion (SLID) bonding is an attractive low-cost wafer-level bonding technology for rapid manufacturing of full Cu3Sn IMC mu bumps. However, SLID requires melting of Sn during the bonding process which poses risks and design challenges in manufacturing. Due to Sn squeeze-out during the bonding process, Sn melt could react with redistribution layers (RDLs) or metallization layers and form IMCs at undesired locations resulting in early failures thereby compromising the reliability. The Sn-squeeze out issue during bonding is addressed in this work by designing test structures with equal and unequal lateral dimensions of mu bumps in the top and bottom wafers. The effects of Sn-squeeze out on the EM resistance and reliability are compared in both designs. Significant improvement in the Sn-squeeze out and corresponding EM resistance was observed in the test structures manufactured with unequal lateral dimensions of mu bumps in the top and bottom wafers. FE element simulations were carried out to gain insights and assess the impact of Sn squeeze-out on the reliability and functionality of the Cu3Sn mu bumps.

Non-destructive inspection and analysis techniques are crucial for quality assessment and defect analysis in various industries. They enable for screening and monitoring of parts and products without alteration or impact, facilitating the exploration of material interactions and defect formation. With increasing complexity in microelectronic technologies, high reliability, robustness and thus, successful failure analysis is essential. Machine learning (ML) approaches have been developed and evaluated for the analysis of acoustic echo signals and time-resolved thermal responses for assessing their ability for defect detection. In the present paper different ML architectures were evaluated, including 1D and 2D convolutional neural networks (CNNs) after transforming time domain data into the spectral- and wavelet domains. Results showed that 2D CNNs processing data in wavelet domain representation performed best, however at the expense of additional computational effort. Furthermore, ML-based analysis was explored for lock-in thermography to detect and locate defects in the axial dimension based on thermal emissions. While promising, further research is needed to fully realize its potential.

Abstract Presentation slides for the ISTFA 2024 Tutorial session “Machine Learning in Semiconductor Failure Analysis: Techniques and Case Studies.”

Abstract This chapter assesses the potential impact of neural networks on package-level failure analysis, the challenges presented by next-generation semiconductor packages, and the measures that can be taken to maximize FA equipment uptime and throughput. It presents examples showing how neural networks have been trained to detect and classify PCB defects, improve signal-to-noise ratios in SEM images, recognize wafer failure patterns, and predict failure modes. It explains how new packaging strategies, particularly stacking and disintegration, complicate fault isolation and evaluates the ability of various imaging methods to locate defects in die stacks. It also presents best practices for sample preparation, inspection, and navigation and offers suggestions for improving the reliability and service life of tools.