Speaker

Ryo Maezono

Japan Advanced Institute of Science and Technology

Date＆Time

2023.02.27（Mon）AM 09:00

Location

吉林大学中心校区唐敖庆楼B区521报告厅

Reporter

        Dr. Maezono is a Professor at JAIST, working on Simulation Science. He got his BSc and PhD at Tokyo Univ. (Applied Physics/condensed matter theory on phase diagrams of magnetic oxides). He was a postdoc at Cavendish Lab., Cambridge Univ. (EPSRC fellow/2000-2002), and moved to NIMS (National Institute of Materials Science, Japan), as a tenure researcher (2001-2007), then moved to JAIST. Since his postdoc in Cambridge, he has worked on ab initio Diffusion Monte Carlo (DMC) calculations using massive parallel computations. As an expert of DMC method, he has given several lectures on many-body problems at several external National Universities in Japan. As a computer scientist, he has contributed also to the education of simulation, which contents are published in his books (ISBN:978-4627818217, 978-4627170315). He also leads several industrial collaborations with companies (Toyota-Motor/Sumitomo-Mining etc.), as well as those with experimental synthesis community in inorganic Chemistry.

Abstract

        The materials characterization using spectroscopy is significantly influenced by major peaks. Whether all the peaks are required to characterize a crystal's structure, or only a specific bandwidth of the spectrum is sufficient would be a fundamental question in the characterization. In this context, we developed a scheme that can identify which peaks are relevant and to what extent by using an auto-encoder to construct a feature space for XRD (X-ray diffraction) peak patterns. Individual XRD patterns are projected onto a single point in a two-dimensional feature space constructed using the encoder. If a point is significantly shifted when a peak of interest is masked, then the peak is relevant for characterizing the pattern represented by the point on the space. In this manner, we can quantitatively formulate the relevance of a peak. By using this scheme, we found peaks with significant intensity, but low relevance help characterize XRD patterns. The irrelevance of a peak cannot be easily explained using physical concepts such as higher-order plane index peaks; it is a machine-learning heuristic finding.