Neural Networks for Quantifying Laboratory Confocal Micro X-ray Fluorescence Measurements

Abstract

The quantification of confocal micro X-ray fluorescence spectroscopy (CMXRF) data obtained with polychromatic excitation in a laboratory setup is challenging. Complex dependencies, an elaborate setup calibration and nontrivial data evaluation makes it a time-consuming and intricate task. In this work we introduce the first application of a neural network for the quantification of homogeneous bulk samples, which significantly simplifies the evaluation and effectively eliminates the need for human input. The training of the neural network is performed on simulated data. For this, a simulation routine for CMXRF data of homogeneous bulk samples is introduced. The neural network is trained to simultaneously quantify the elemental concentrations of 53 elements, the density of the sample and the surface position directly from depth profiling measurements. As a result, the CMXRF evaluation is substantially simplified and the potential of the used neural network for feature extraction and prediction is demonstrated.