edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 1 6 tional, machine learning or deep learning methods. The comparison of experimental and reference volume may be made on the 2D slices or directly on the 3D structures. This article explores reference-based data processing with conventional slice-by-slice comparison and with deep learning. Comparing the experimental dataset with a second, defect-free one would be excessively time-consuming, as the sample region investigated is a different one in most cases. Instead, 2D CAD layout images of the same volume are extracted and used as ground truth, to detect defects by searching for structural differences between the expected (CAD) and the measured data (Fig. 5). Slice-to-slice comparison of datasets requires matching slices with identical structural content, excluding defects. Mismatches in position, scale, or skew can be misinterpreted as defect signatures. Affine registration can align datasets in 3D but is insufficient if real device structures deviate from their layout in scale, shape, and position. Consequently, elastic registration using optical flow was employed for better adaptation of CAD to experimental. Careful adjustment of registration parameters is crucial to prevent deforming reference structures to mimic defect shapes. After registration, a 3D median filter (2 pixel) was applied to the experimental volume to reduce noise. Following binarization by gray value thresholding, spurious noise up to 3 pixels was removed. Finally, 2-pixel erosion was performed to reduce structure sizes to CAD structure widths. Defect signature detection was explored using three gray value-based methods: A) Comparing mean gray values within each slice pair; B) Calculating cross-correlations between each slice pair. Both A and B were applied to the XY, YZ, and XZ slices; C) Performing image arithmetic with the CAD data mask-ing structures in the experimental data. Method A, shown in Fig. 6, plots mean gray values for XZ, XY, and YZ slice pairs against slice number. Within the expected slice ranges, no significant difference is observed between experimental and reference. This is consistent with cross-correlation analysis (method B), where Fig. 7 shows normalized cross-correlation error over slice number without deviations attributable to the defect. The reason for the lack of a defect signa- ture with these methods becomes clear look- ing at the difference between two slices. Figure 8 shows the difference between an XZ slice from the FIB-SEM dataset extracted at the M1 layer containing the defect, and its CAD counterpart. It is evident that despite extensive post-processing, the contribution of the defect to the gray value distribution is small. Real structures not perfectly matching reference structures and residual image registration errors outweigh the effect of the defect on mean gray values. Method C continued with the difference between binarized measured data and CAD (Figs. 9a-d). To get rid of ghost structures due to deviations between experimental and layout data, it was assumed that the CAD ground truth represents a Fig. 7 Normalized cross-correlation error between each tomography slice of the FIB-SEM and the corresponding CAD slice, plotted over slice number, for XY, YZ, and XZ slices. Gray rectangles highlight slice ranges with the defect.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==