edfas.org 3 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 1 EDFAAO (2026) 1:3-10 1537-0755/$19.00 ©ASM International® A STEP TOWARD AUTOMATION IN FAILURE ANALYSIS BY FIB-SEM 3D TOMOGRAPHY AND AI SEGMENTATION Pascal Limbecker1, Rong Wu1, Thomas Woyack1, Heiko Stegmann2, Daniel Plencner2, and Roland Salzer2 1GlobalFoundries LLC & Co. KG, Dresden, Germany 2Carl Zeiss Microscopy GmbH, Munich, Germany pascal.limbecker@globalfoundries.com INTRODUCTION In the competitive high-volume semiconductor foundry business, yield, quality, and cost are key to gaining competitive advantage. In addition to in-line metrology and defect scanning, mass failure analysis of base line product wafers is important to generate defect Pareto charts for continuous yield improvement, by eliminating the major yield detractors. Hundreds of short or leakage failures, localized by optical beam induced resistance change (OBIRCH) or photon emission, plus several thousands of bitmap and logic failures are analyzed using FIB-SEM slice and view or top-down polishing and SEM inspection every year, requiring lots of manpower and tool time. Over the past decades, visitors to FA laboratories have often asked operators if they could imagine a computer taking over their job. The typical answer was: “No, you need the human eye and mind to detect a defect.” However, the evolution of automation, artificial intelligence, deep learning, and improved FIB-SEM performance has changed the mindset. Operators can navigate to the area of interest using CAD alignment on full wafers, wafer pieces, or single dies. Enabling bitmapping (memory) and scan diagnosis (logic) narrows down the failure location to a small area for slice and view application. OBIRCH, photon emission, or lock-in thermography also provide a precise failure location. Because top-down inspection by delayering multiple dies is very time consuming, further physical analysis is done by manual FIB slice and view. The maximum area to be investigated is typically not larger than 15 × 15 µm2, and is usually smaller. The direction of the cut must be decided because it is not known in which layer the defect is located and there is a risk that the defect will be missed in the selected direction.[1] AUTOMATED FIB-SEM SLICE AND VIEW Recent FIB-SEM improvements allow for acquiring a series of images of the target volume fully automatically. Using such FIB-SEM tomography data for root cause analysis at defect sites isolated by volume diagnosis techniques has been demonstrated.[1] In the case study of logic failure presented here, the failure area was reduced to 10 × 10 µm² through scan diagnosis. A FIB-SEM tomography dataset was acquired in a ZEISS Crossbeam 550L FIB-SEM, using Atlas 3D software, which is capable of autofocus, autostigmation, and ROI centering using special fiducial patterns[2] in an ion beam deposited platinum and carbon layer sandwich (Fig. 1). The data acquisition extended over the entire area of 10 × 10 µm2 to 5 µm depth. Nominal FIB slice thickness was set to 5 nm. High resolution SEM imaging of each slice was done at 5 nm pixel size on a 10 × 1 µm2 sub-area covering the complete layer stack. Both secondary electron (SE) and backscattered electron (BSE) data were stored simultaneously. Slice thickness variations and lateral drifts during data acquisition can lead to distortions of the sample structures in the dataset. To counteract, the fiducials serve as references for periodic slice thickness and drift correction.[2] Data acquisition took 15:30 h. Traditional slice and view inspection of the defect would have taken an estimated 2 to Fig. 1 Fiducial patterns on the top region of interest.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==