Over the past decade, the semiconductor industry has increasingly focused on packaging innovations to improve device performance, power efficiency, and reduce manufacturing cost. The recent heterogeneous integration offers an attractive solution in advanced IC packaging because it enables the integration of diverse functional components, such as logic, memory, power modulator, sensor on a single package platform. However, the adoption of the emerging structures, materials and components in advanced packages has challenged the existing fault isolation and analysis techniques. One of the major challenges is the limited accessibility to defects because fault regions are often located deep within devices. Without high-accuracy positional information of a defect, physical cross-sectioning and FIB polishing may alter or destroy the evidence of root causes. A non-destructive microscopic approach is preferred to map defective sites and surrounding structures. However, this method is limited by spatial resolution, especially for analyzing novel submicron interconnects such as fine pitch microbumps, redistribution layers (RDLs), and hybrid bonds. In this paper, we report an AI powered correlative microscopic workflow, where non-destructive X-ray imaging, FIB polishing and high-resolution SEM analyzing techniques are combined to solve the accessibility problem. Because 3D X-ray imaging may take a larger fraction of the time span over the entire workflow, a deep-learning based reconstruction method was applied to accelerate data acquisition. Several next-generation packages, fan-out wafer-level package (FOWLP) and hybrid bonds with sub 10 µm pitch, were used as the test vehicles to demonstrate the workflow performance and efficiency.

In this work we present a new defect localization capability on Wafer Level Chip Scale Packages (WLCSP) with small-scale Cu pillars using advanced 3D X-ray microscopy (XRM). In comparison to conventional microcomputed tomography (Micro-CT or microCT) flat-panel technology, the synchrotron-based optically enhanced 3D X-ray microscopy can detect very small defects with submicron resolutions. Two case studies on actual failures (one from the assembly process and one from reliability testing) will be discussed to demonstrate this powerful defect localization technique. Using the tool has helped speed up the failure analysis (FA) process by locating the defects non-destructively in a matter of hours instead of days or weeks as needed with destructive physical failure analysis.

3D X-ray microscopy (XRM) is an effective highresolution and non-destructive tool for semiconductor package level failure analysis. One limitation with XRM is the ability to achieve high-resolution 3D images over large fields of view (FOVs) within acceptable scan times. As modern semiconductor packages become more complex, there are increasing demands for 3D X-ray instruments to image encapsulated structures and failures with high productivity and efficiency. With the challenge to precisely localize fault regions, it may require high-resolution imaging with a FOV of tens of millimeters. This may take over hundreds of hours of scans if many high-resolution but small-volume scans are performed and followed with the conventional 3D registration and stitches. In this work, a novel deep learning reconstruction method and workflow to address the issue of achieving highresolution imaging over a large FOV is reported. The AI powered technique and workflow can be used to restore the resolution over the large FOV scan with only a high-resolution and a large FOV scan. Additionally, the 3D registration and stitch workflow are automated to achieve the large FOV images with a recovered resolution comparable to the actual high-resolution scan.

Microscopic imaging and characterization of semiconductor devices and material properties often begin with a sample preparation step. A variety of sample preparation methods such as mechanical lapping and broad ion beam (BIB) milling have been widely used in physical failure analysis (FPA) workflows, allowing internal defects to be analyzed with high-resolution scanning electron microscopy (SEM). However, these traditional methods become less effective for more complicated semiconductor devices, because the cross-sectioning accuracy and reliability do not satisfy the need to inspect nanometer scale structures. Recent trends on multi-chip stacking and heterogenous integration exacerbate the ineffectiveness. Additionally, the surface prepared by these methods are not sufficient for high-resolution imaging, often resulting in distorted sample information. In this work, we report a novel correlative workflow to improve the cross-sectioning accuracy and generate distortion-free surface for SEM analysis. Several semiconductor samples were imaged with 3D X-ray microscopy (XRM) in a non-destructive manner, yielding volumetric data for users to visualize and navigate at submicron accuracy in three dimensions. With the XRM data to serve as 3D maps of true package structures, the possibility to miss or destroy the fault regions is largely eliminated in PFA workflows. In addition to the correlative workflow, we will also demonstrate a proprietary micromachining process which is capable of preparing deformation-free surfaces for SEM analysis.

3D X-ray tomography plays a critical role in electronic device failure analysis, but it can take several hours to overnight to get sufficient resolution in fault regions to detect and identify defects. In this paper, we propose a machine learning based reconstruction technique that can speed up data acquisition by a factor of four or more, while maintaining image quality. The method, which uses neural networks, extracts signals from low-dose data more efficiently than the conventional Feldkamp-Davis-Kress (FDK) approach, which is sensitive to noise and prone to aliasing errors. Several semiconductor packages and a commercial smartwatch battery module are analyzed using the new technique and the results compared with those obtained using conventional methods. The neural network can be trained on as little as one tomography image and the only requirement for the training data is that the sample or region of interest is well represented with all characteristic features in the field-of-view.