edfas.org 5 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 1 compared to manual slice and view. However, total acquisition time is determined by the ratio of voxel size to volume imaged and can be reduced by taking more information about the defect into consideration: The more precisely the defect location is known, the more the acquisition volume can be reduced. If the defect size is roughly known, the voxel size can be increased to minimize acquisition time while maintaining sufficient 3D resolution. In hindsight it is estimated that the acquisition time may have been reduced to below 3 h without compromising the information content of the data. AUTOMATED DEFECT DETECTION The next step toward automation is automated defect detection. For 2D electron microscopic images, automatic defect finding is a current field of research, but the authors are not aware of existing work on automatic defect localization in 3D FIB-SEM tomography data of semiconductor devices.[6] Various aspects and challenges of the topic are highlighted here. Finding a defect of known nature, but unknown shape and size in a complex 3D interconnect structure is more challenging than typical 2D applications, such as the classification of different defect categories in SEM images of interconnect line arrays.[7] As in the 2D case, reference-based and reference-free methods can be consid- ered, further subdivided into conven- Fig. 4 3D rendering of AI-based segmentation of M1 layer. In the center, the defect is visible. Fig. 5 Top: cross-section SEM image. Bottom: CAD image from same location. Fig. 6 Method A: Mean gray values of each image in the tomography (red) and CAD datasets (blue), plotted over slice number, for XY, YZ, and XZ slices. Gray rectangles highlight slice ranges with the defect.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==