Publikationsrepositorium - Helmholtz-Zentrum Dresden-Rossendorf

To extract chemical compositions and layer thicknesses of layered samples from back scattering spectra experimentalists usually have to take the following approach: Simulation of a theoretical spectrum for an initial target configuration and comparison to the measured data followed by the successive re-adjustment of the target model iteratively, until simulation result and experimental spectrum fit together. For multi-layer samples this procedure can get rather time consuming, especially when a series of similar samples with varying layer thickness and/or stoichiometry has to be analyzed.

Although modern IBA spectrum simulation software like SimNRA[1] or WINDF[2] have become quite powerful and handy tools, the analysis of the spectra consumes still a significant fraction of an IBA scientist’s working time. SimNRA offers therefore the opportunity to partially fit layer thicknesses and/or elemental ratios for a given layer within a certain region of a spectrum. WINDF goes a step further and implements an automated spectrum fitting based on a simulated annealing algorithm. However, it takes the user quite some time to set up the boundary conditions and fit parameters until the actual fit procedure can be initiated. Furthermore, the outcome of the fit procedure in some cases contains non-physical artifacts and requires multiple re-adjustments of the boundary conditions / fit parameters.

An approach that came up in the past (and is still being applied for particular tasks) is the application of artificial neural networks (ANN) to derive sample information from IBA spectra [3,4]. In a nut-shell this method basically trains an algorithm how the shape of a spectrum is correlated to the sample’s target model without introducing any physics (numerical calculations) to the code. Therefore, the ANN is fed with many (typically several 10 thousand) training spectra with a known target model. After this training procedure (which can be quite time consuming) the ANN spits out the target model of any unknown spectra in almost zero time. However, the spectra must be of the same type as all the training spectra since an ANN can only interpolate and not extrapolate, which is for sure one of the mayor drawbacks of this approach. However, all these efforts are justified in some special scenarios e.g. if a large series of spectra of similar type has to be evaluated.

In this contribution, we present a new approach of automated IBA spectra fitting applying an evolutionary algorithm (EA). We show that EA is well suited and robust for complete and fast IBA spectrum fitting with minimum input of boundary conditions. The benefits of this algorithm and the particular differences to simulated annealing and ANN are pointed out. Special emphasis is put on the adoption of this algorithm to the analysis of MEIS spectra, since there is a couple of differences to classical IBA methods that needs to be considered.

Based on this algorithm a platform independent software package has been developed that comprises a clean and easy-to-use graphical user interface. We will introduce this software in a basic overview.