Optimal Design of Experiment for X-Ray Spectromicroscopy by Machine Learning

The total measurement time of an X-ray spectromicroscopy experiment using a scanning transmission X-ray microscope (STXM) is determined by a multiplication of a number of energy points ne, sample scanning points ns, and measurement time per each point tm plus overhead. Overhead consists of time for data acquisition, moving of sample scanners, beamline optics and undulator properties (gap and phase of magnet arrays). An X-ray spectromicroscopy experiment with an STXM is performed as an image acquisition by sample scanning in an energy-by-energy regime. Moreover, moving of beamline optics such as a grating and mirrors takes longer time than that of piezoelectric actuators for sample scanning. Therefore, it is a good strategy to reduce ne to reduce total measurement time. Another strategy to reduce total measurement time is an optimization of tm. One can reduce tm at the expense of a signal-to-noise (S/N) ratio of spectra, which is proportional to tm1/2. It is important to reduce total measurement time without degrading the quality of spectra to extract physical or chemical parameters by analysis. Machine learning techniques are expected to resolve this issue. Ueno et al. proposed the adaptive design of an X- ray magnetic circular dichroism (XMCD) spectroscopy experiment by Gaussian process (GP) modeling, a machine learning technique, and successfully reduced the total number of energy points to measure to evaluate magnetic moments with required accuracy