Virtual Experiments combined with Machine Learning to improve Data Evaluation of SANS Measurements

Topic  4
Main supervisor K.Lieutenant (k.lieutenant@fz-juelich.de)
MLZ institution FZJ
Local supervisor 1 P.Guillot
Institution Solvay
Local supervisor 2
Institution
Local supervisor 3
Institution
Local supervisor 4
Institution
Title Virtual Experiments combined with Machine Learning to improve Data Evaluation of SANS Measurements
Description With the growing importance of nanotechnology, SANS measurements are performed by more and more scientists. However, getting information from the SANS spectrum remains difficult, as there are only few characteristics features that can be related to sample properties. Following the idea promoted by the Helmholtz Society to improve data evaluation by artificial intelligence, we want to combine virtual experiments with machine learning to improve data evaluation. The advantage compared to real experiments is that the sample properties can be precisely defined and that (using a computer cluster) a high number of experiments can be performed in a relatively short time with low costs. The advantage compared to using tools for spectra simulation is, that characteristics of the instrument can be included in the simulation. The virtual experiments would be used to train the program for evaluation of SANS data. Later on, the program can be used to help interpreting real data. The project comprises the following tasks: – checking and improving the VITESS module ‘sample_SANS’ that simulates small-angle scattering; – choosing a real SANS instrument and simulating it using VITESS; – getting perfect agreement between real and virtual experiment for a few selected samples; – combining the simulation with the data evaluation used for the real data; – combining the virtual experiment and its data evaluation with a suited machine learning program; – training this program by varying size, shape and distribution of the scattering particles systematically; – checking the program on known samples, first in virtual and then in real experiments. The project aims to explore the potential of machine learning for SANS data, when the software is faced with data from unknown samples.