The structure of materials can be mapped in 3D at the nanoscale by the FIB-SEM serial sectioning technique. However, in the case of high porosity, the structure that can be directly reconstructed from the 2D FIB sections does not correspond to the real 3D structure, since structural regions behind the current section surface are also visible in the SEM images due to the pores. "Classical" solutions to this segmentation problem are typically tailored for one material and fixed imaging parameters.
Machine and statistical learning techniques have higher potential for a generalized solution. However, training data are difficult to obtain. Synthetic FIB-SEM images, for which the ground truth is known, provide an attractive way out. Synthetic FIB-SEM images of realizations of random sphere and cylinder systems can be used to train a CNN with U-net3d architecture for the segmentation task. However, it must be trained again with a very similar structure for new structures. In particular, quantum computing simulation and learning methods are also being researched, developed, and tested.