3D characterization at the nanoscale by stereoscopic scanning electron microscopy

Scanning Electron Microscopy (SEM) is a 2D microscopy technique, providing information related to morphology, composition and other properties of bulky samples. We discuss the access to 3D characterization down to the nanoscale based on stereoscopy, a technique already well assessed on the macroscopic scale. In the simplest case, as in the human vision, at least a couple of stereo images are acquired at different inclination. The relative displacement of a feature in the different images allows to determine its elevation with respect to a reference plane. In this work, the process was tested on images showing details at the scale of tens of nm. The position of sparse points at sample surface was reconstructed according to several approaches. First, the features in the two images were directly recognized and matched by the operator, in order to maintain the highest degree of control on the process. A 3D positional uncertainty lower than 14 nm was obtained for small nanostructures lying on regular surfaces, evaluated in excess as the standard deviation of the set of reconstructed positions from the geometrical reference surface. The main limiting factors for the 3D reconstruction result to be the limited amount of uniquely recognizable features, the ratio between field of view and resolution and the contrast attainable at high magnification. As a second step, several strategies for the automatic feature recognition on each image and for their matching between different image projections were tested. The most promising algorithm proved to be the so called Speeded Up Robust Features local detector and descriptor. A rather conservative set of parameters was used in order to limit the results only to reliably matched features. The algorithm proved to be able to position the nanostructure center of mass with a reliability of the order of 2 nm, comparable to lateral resolution of the SEM images. The automatic approach could result in a reduction, by several order of magnitudes, of the number of matched features with respect to the direct operator recognition, depending on the choice of parameters and filters. Nevertheless, the 3D standard deviation from the reference geometrical surface found in this case was estimated lower than 12 nm. Fully automatic reconstruction may be a reliable option only when the feature contrast is outstanding and their distribution dense. Similar reconstructions were done on dense distributions of 50 nm sized structures on silver ink drops deposed on spin coated polymer substrates. These show rougher surfaces where a standard deviation of 18 nm from the reference geometrical surface was obtained.