Speaker
Description
The data obtained from the synchrotron generally requires a considerable amount of time to be analyzed (months or years). As part of this project (funded by the PEPR DIADEM and started in February 2025), we aim to apply neural network algorithms in order to obtain a first estimate of the results directly from the raw data, in real time during the experiment.
In the first phase of the project, we decided to select a simple technique that is common to most beamlines. Moreover, since neural networks require thousands of labeled data samples, it was necessary to choose a technique that can be easily simulated.
The choice fell on X-ray fluorescence (XRF); this technique can be simulated with a high degree of accuracy using software called McXtrace. The idea is to determine the stoichiometry of the measured material directly from the raw spectrum, thereby avoiding manual analyses.
In this presentation, we will describe the algorithm developed during this initial phase of the project. The algorithm can already predict the stoichiometry of 51 elements from XRF spectra for synthetic data, with a variable photon energy and detector energy (between 3 and 60 keV). The results on simulated data are promising, with a coefficient of determination of R² = 94%.
The next step is to apply the model to real XRF data. To achieve this, we plan to fine-tune the model parameters using the experimental spectra. Furthermore, the approach will be extended to other techniques such as XRD or XAS.
