edfas.org 7 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 1 minimum subset of the experimental data in terms of structure sizes. In other words, wherever the CAD data shows structure, a corresponding structure is found in the measured data, unless there is a defect of type “missing structure.” Hence, where the CAD is not predicting anything, the data can be ignored. This was achieved by multiplication of each difference image with the corresponding CAD image. The result is shown in Fig. 9e. After dilation and erosion to remove remaining small artifacts and noise, the defect is segmented (Fig. 9f). Figure 9g shows an overlay of the experimental image with the defect mask. This simple approach provides good results. The reconstructed XY-view shows the found defect areas marked in red (Fig. 10). Some false positives are visible, but it would be easy to ignore them in a subsequent automation step. Note that this algorithm only works for the defect class “opens,” and not for additional features not in the CAD, such as particles. For further automation of the method, feedback loops using AI models are a promising approach for better registration. (a) (b) (c) (d) (e) (f) (g) Fig. 8 Difference between two corresponding XZ slices from the M1 layer of both datasets, showing the defect (compare with Fig. 2). Fig. 9 Image processing steps to detect a defect in the tomography image stack. (a) raw, (b) binarized BSE image, (c) CAD slice, (d) difference between (b) and (c), (e) shows (d) multiplied with (c), (f) shows (e) after erosion and dilation, and (g) overlay of (a) (aqua) and (f) (red). Fig. 11 Normalized aspect ratios of experimental (red circles) and CAD layout segments (blue squares) plotted against their normalized volume. Two outlier points marked by a red circle belong to the two halves of the open interconnect. Fig. 10 Detail of a virtual top-down image of first metal layer showing a defect in the center, extracted from the post-processed dataset in XZ.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==