edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 1 4 3 hours. However, unattended tomography data acquisition maximizes tool utilization. DATASET POST—PROCESSING AND VISUALIZATION After acquisition, all sequential images must be registered perfectly to create a precise 3D model. Translational alignment by cross-correlation to remove residual image to image drifts was applied. Using the slice thicknesses measured during acquisition, a constant slice thickness of 5 nm was interpolated. The final dataset consisted of 1702 SE and 1702 BSE images of 1987 × 408 pixels each. The quality of the dataset can be assessed in all directions from virtual slices extracted in XZ and YZ directions and is impressive. The virtual top-down image in XZ direction (Fig. 2) showing a pattern defect in M1 can hardly be distinguished from a top-down image acquired after delayering by manual polishing or PFIB. For better data interpretation by process engineers, a 3D visualization is helpful. A prerequisite for such visualization is a segmentation of individual structures. This is straightforward with artificial intelligence-based segmentation.[3] It consists of defining a number of segmentation classes, then annotating a limited number of training images according to these classes, training the model on these annotations, and finally using the model to segment the full dataset or other datasets acquired with the same conditions. Ten classes were annotated in 25 tomography images: metal layers M1 to M5, Si substrate, contacts, NiSi and poly-Si; everything else as background. These annotations were then used to train a segmentation model in 1200 epochs. Annotation and model training was done with the ZEISS Arivis Cloud service. It uses a U-net with EfficientNet B0 as encoder and Pixelshuffle as decoder.[4,5] The learning rate was 3e-4 for the first 70% of epochs and 4e-5 for the remainder. Batch size was 3. The algorithm returns several quality measures, intersection over union (IoU) was specified as 0.8. For segmentation of the full dataset with the model and for visualization, ZEISS Arivis Pro desktop software was used. The result contained spurious small, artifactual segments. They were removed by filtering out segments with unrealistically small volumes. A volumetric rendering of the dataset and of the segments is shown in Figs. 3a and b. The segmented M1 layer with the defect is shown in Fig. 4. Applying the model to another dataset obtained from a different sample with similar open interconnect defect produced too many artifacts, indicating that the model did not yet generalize well enough to be applicable to different datasets. In this case study, data acquisition time was long (a) (b) Fig. 2 Virtual XZ top-down image of first metal layer showing a defect in the center marked by a red oval, extracted from the post-processed BSE dataset. Fig. 3 Volumetric 3D renderings of (a) BSE dataset and (b) AI-based. Si substrate not shown.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==