Speaker
Description
In modern material science the amount of generated experimental data is rapidly increasing while analysis methods still require many manual work hours. Especially, this is the case for X-ray photoelectron spectroscopy (XPS), where quantification is a complex task and, in many cases, can be properly done by experts only. However, these problems could be overcome by the use of a neural network-based approach (NNA). An important question is to validate NNA-based results and to compare them with results obtained with a use of a commonly used, manual fitting procedure. Since the available experimental data is insufficient for network training, a synthetic dataset was created using parameters obtained from the real experimental XP spectra measured on a reference sample. A 4-component model has been chosen and some parameters (binding energies, FWHM) were almost fixed, but the peak intensity was a free parameter keeping the total area of all 4 components as constant (normalization). As commonly used fitting procedure CasaXPS software was applied. After training on the synthetic data, the neural network was tested on the experimental data obtained from another sample and predicted area percentages were compared with the fitting results. The predicted area percentages are in good agreement with corresponding area percentages from the fitting with CasaXPS. Moreover, it was established that it is crucial to choose a proper model and corresponding NNA training set, otherwise the experimental data could not be evaluated properly. It means that this approach can therefore be successfully used not only for XPS quantification tasks directly, but also to validate proposed models.