Logo of the Physikalisch-Technische Bundesanstalt

Machine learning for reconstruction of periodic nanostructures


The characterization of nanostructured surfaces with a sensitivity in the sub-nm range is of great importance for the development of the next generation of integrated electronic circuits. At PTB, various X-ray techniques have been further developed in recent years, including the grazing incidence X-ray fluorescence (GIXRF) method. In this technique, a standing wave field is formed all around the nanostructures, and the emitted X-ray fluorescence is analyzed spatially and energetically at different angles of incidence. However, the possible reconstruction of the nanostructure from this effect has so far required an enormous numerical effort due to the necessary optimization processes. Thus, an analysis of uncertainties based on modern statistical methods has not been possible. However, the use of machine learning opens up completely new approaches in the field of nanometrology: For example, it has now been possible for the first time to reconstruct a Si3N4 grating structure with high precision in an element-sensitive manner using GIXRF techniques. A so-called Bayesian optimization was used, which not only considerably accelerates the optimization process by means of intelligent search strategies. At the same time, it also allows initial estimation of the possible uncertainties that arise in the parameterizations of the nanostructures and the underlying simplified physical models. 


A. Andrle et al., Nanomaterials 2021, 11(7), 1647 https://doi.org/10.3390/nano11071647


A. Andrle, 7.14, E-Mail: Opens local program for sending emailAnna.Andrle(at)ptb.de