A Machine Learning Approach on Analysis of Emission Spectra for Application in XFEL Experiments

Abstract: In this thesis we investigate two potential applications of machine learning in the context of X-ray imaging and spectroscopy of biological samples, particularly such using X-ray free electron lasers (XFEL). We first investigate the possibility of using an emission spectrum, recorded from a sample after being probed by an incident X-ray, as a diagnostic tool. We produced a training dataset of simulated emission spectra, where the incident X-ray energy and fluence was varied as well as the sample density. The simulations were implemented using Cretin which is a radiation transfer code which model the behaviour of plasma. We then trained a dense neural network to predict the three above named features given an emission spectrum. The dependency between input and output is inherently non-linear, making neural networks a suitable method for these predictions. Our results show a mean prediction error of below 6% of the entire range of all three features. If a similar tool was to be implemented in real life XFEL experiments, it could provide useful information in the data analysis pipeline.   As a second focus of this thesis we aim to produce an application to be used by researchers in XFEL experiments. Given a set of input parameters, including the incident X-ray energy and fluence along with atomic content and density of the sample, our application generates an emission spectrum for the user. The application is based on a neural network trained on Cretin simulations. When evaluated by comparing the final model to simulations, our model was found to have a mean absolute percentage prediction error of 1.77%. In addition to this we include similar models that generate the time development of the electron temperature and mean ionization of the sample, since these properties are highly associated with the emission processes of plasma. We did this by training dense neural networks on a dataset consisting of simulations of the corresponding property. Finally we integrated our models in a graphical user interface web application, accessible via the QR code. With this approach, the desired data can be plotted in real-time in a user-friendly manner, without having to run complicated and time-consuming simulations. Our model is focused on biological samples and could be used as a reference tool in structural biology.