Jun 14 – 18, 2021
Europe/Stockholm timezone

Artificial Intelligence Analysis of Reflectivity Data

Jun 18, 2021, 8:00 PM




Stefan Kowarik


Stefan Kowarik1, Alessandro Greco2, David Marecek1, Erich Hüthmair1, Vladimir Starostin2, Alexander Hinderhofer2, Alexander Gerlach2, Maximilian Skoda3, and Frank Schreiber2
— 1Department of Physical Chemistry, University of Graz, Austria
— 2Institute of Applied Physics, University of Tübingen, Germany
— 3Rutherford Appleton Lab, ISIS Neutron and Muon Source, UK
We review the applicability of artificial neural networks to analyse X-ray reflectivity (XRR) and neutron reflectivity data. Compared to standard iterative fitting approaches, the neural network analysis can predict sample parameters without needing a good initial guess of the fit parameters. Also, a neural network analysis copes well with a low signal to noise ratio and a high background signal as long as features such as the total reflection edge are prominent. However, the prediction accuracy is lower than a standard fit with parameter errors around 10 %.
In comparison with iterative fitting the neural network analysis is orders of magnitude faster, which is beneficial e.g. for estimates of parameter errors via quick predictions of multiple ‘leave-one-out’ / jack-knife XRR datasets. The processing speed also offers advantages for (on the fly) batch processing of large datasets e.g. from real-time experiments. Lastly, neural networks can co-refine multiple reflectivity curves, while taking into account prior information. For the example of XRR curves acquired during thin film growth this reduces parameter errors when compared to individual fits and also enables predictions not only of thickness, roughness or density but also of temporal parameters such as growth rate.
[1] Greco et al., J. Appl. Cryst., 52, 1342 (2019)
[2] Greco et al. Mach. Learn.: Sci. Technol. in press https://doi.org/10.1088/2632-2153/abf9b1 (2021)

Presentation materials

There are no materials yet.