Presenter Profile
Motoki Shiga
Professor
Tohoku University, Unprecedented-scale Data Analytics Center
Motoki Shiga obtained his B.Eng. (2001), M.Eng. (2003), and D.Eng. (2006) degree from Gifu University. He was a postdoctoral researcher (2006–2007) and an assistant professor (2008-2011) in Kyoto University, an assistant professor (2011-2013) in Toyohashi University of Technology, and an assistant professor (2013-2017) in Gifu University. Then, he promoted to an associate professor (2017–2022) in Gifu University. He moved to Tohoku University as a full professor in 2022. His research interest is machine learning and its application for materials science.
TALK TITLE
Material structure analysis based on machine learning
KEYWORDS
Machine learning, Materials informatics, Spectral imaging
ABSTRACT
Spectral image (SI), which is an imaging technique to observe the spectra over the spatial grid points in a region of interest (ROI), is useful for identifying spectra and spatial distribution of chemical components in new materials. There exist several techniques such as canning transmission electron microscopy combined with electron energy-loss spectroscopy or energy-dispersive X-ray spectroscopy, and X-ray absorption fine structure (XAFS) imaging.
Due to the large size of SI data, the analysis cost is a bottleneck in the evaluation process of material structures. Against this problem, machine learning techniques are useful because they can automatically extract essential structural information from the observed massive SI data. This talk introduces our developed machine learning methods based on non-negative matrix factorization (NMF) for SI data. The assumption of basic NMF, which is a linear mixing of nonnegative values of spectra and intensities of chemical component, is simple but effective for analyzes of a lot of applications.
Our research group proposed an extension realized by introducing a spatial orthogonality constraint and a sparsity constraint to obtain physically meaningful results. This talk also introduces another extended method to analyze three-dimensional structure data by computed tomography (CT) techniques. The effectiveness of these methods are demonstrated using real experimental datasets.