Abstract:
X-ray and neutron scattering encompass a large variety of complementary and non-invasive measurement techniques that are used to study a large range of materials. The continued improvements of X-ray and neutron sources, as well as advancements in detector technologies have enabled the development of sophisticated measurement setups with the ability to gather large amounts of information. These modern techniques tend to produce a lot of data, which is typically analyzed using theoretical models. Increasingly, however, the rate at which data is produced is outpacing the rate at which it can be analyzed. To fully exploit the potential of these techniques and avoid bottlenecks in scientific productivity, equally advanced methods of data analysis must be developed. In recent years, machine learning, specifically deep neural networks (NNs), have emerged as a promising solution for this, since their data-driven heuristic models can often process data many times faster than conventional methods.
The research in this work focuses on the first published application of NNs for the analysis of specular reflectometry data and demonstrates further improvements of the method. X-ray and neutron reflectometry are commonly used to study various important systems, such as surfaces, interfaces, liquid and solid thin films, layered structures and magnetic materials. As shown in this work, NNs can extract sample properties from reflectometry data within a fraction of a second, which is on par with the high-end speed of modern measurements. This is demonstrated with, but not limited to, organic molecular thin films on silicon substrates. Furthermore, this work discusses the NN performance on different challenging cases and shows methods of successfully dealing with systematic and statistical artifacts in the data. This thesis culminated in the development of mlreflect, a Python-based analysis package that implements the achievements of this work and that is available both online and on the Maxwell cluster of the Deutsches Elektronen-Synchrotron. Thus, this work constitutes a significant step towards the goal of fully automatized reflectivity data analysis and may even serve as a guide for the analysis of other types of scattering data.