Studie van Germanidevorming door Machine-Learning-gebaseerde RBS-analyse

Developments in ion beam analysis (IBA) methodology, including sophisticated setups and advanced measurement schemes, enable the study of increasingly complex materials, even under restrictive conditions. However, the resulting data are often suboptimal. Additionally, advancements in data evaluation approaches expand the application field of IBA. It facilitates the analysis of complex data from, amongst others, cultural heritage investigations, the study of technologically relevant systems, and scientific research driven by fundamental interest. Our specific interest lies in using IBA for the fundamental study of solid-phase reactions involving thin films, which often yields suboptimal data. In such reactions, the sequence of phase formation may differ from bulk solid-phase reactions due to the reduced length scales. A pertinent example is the (extremely) thin-film reaction between a metal and a Ge(Sn)-based substrate, known as (stano)germanidation, which also has practical applications in the formation of contact materials in transistor technology. Real-time Rutherford backscattering spectrometry (RBS) - i.e., measuring RBS spectra while thermally annealing the sample - provides a systematic and high-temperature resolution investigation into the growth and consumption of compounds during such solid-phase reactions. Nevertheless, it typically yields suboptimal data sets due to reduced measurement times, resulting in low-statistics data. Moreover, the RBS analysis of these multilayered structures consisting of extremely thin films results in highly convoluted spectra that are challenging and possibly ambiguous to evaluate. To address these challenges in analyzing complex RBS data, backscattered ions can be collected in multiple scattering geometries followed by a simultaneous evaluation of the collected spectra, also known as self-consistent evaluation. This allows to reduce or even avoid ambiguities in the data analysis. In this thesis, we have developed a self-consistent evaluation approach for complex RBS data based on artificial neural networks (ANN). We extensively test this approach and investigate how self-consistent evaluation impacts the combined analysis uncertainty and the analysis reliability, even in the case of data collection with inaccurately known setup parameters. The focus of our ANN development and testing is on meeting the specific analysis requirements for real-time RBS data, particularly in probing the formation of (stano)germanides. ANN analysis is employed to classify the RBS spectra with a reduced potential user bias, which allows the identification of annealing temperature ranges in which specific (stano)germanide phases grow by consuming previously formed phases. Thereby, ANN classification of RBS data establishes the sequence of phase formation of the metal (stano)germanides. Our focus is on a detailed analysis of (stano)germanidation involving the solid-phase reaction of Pt with various Ge(Sn)-based substrates. The evaluation of the highly complex data sets using the developed ANN approach is essential to obtain an accurate and precise analysis of the Pt (and Sn) redistribution. From the resulting compositional depth profiles of the (stano)germanides as a function of annealing temperature, we explore the kinetic and thermodynamic driving forces behind phase growth through nucleation and diffusion. Furthermore, we investigate changes in the solid-phase reaction when modifying the single-crystal Ge(001) substrate. Special attention is devoted to how the onset temperature of phase growth changes and how the stability of the Ge-rich phases over an extended temperature range is affected by these modifications.