Kiyou Shibata1,Teruyasu Mizoguchi1
The University of Tokyo1
Kiyou Shibata1,Teruyasu Mizoguchi1
The University of Tokyo1
Recently, in the context of materials informatics, there have been many attempts to replace expensive first-principles calculations with predictions based on low-cost machine learning (ML) models. Although databases of structures and electronic states of various materials are already being developed and play an important role in the study of ML models, it is necessary to expand not only the number of data in the databases but also their contents in order to construct ML models that represent a wide variety of physical properties.<br/>Among various physical properties, we are focusing on the core-loss spectra, which are not only computationally available but also experimentally measurable by transmission electron microscopy and synchrotron facilities. The core-loss spectra, which are excitation spectra from a core orbital to unoccupied orbitals, directly reflect partial electronic state densities (PDOS) of states excited state and indirectly contain information on ground-state PDOS, and thus on properties related to PDOS. In the context of these backgrounds, we have been performing first-principles calculations on the K-edge core-loss spectra of each independent carbon site in the organic molecules in the QM9 data set [1], which have carbon numbers of 8 or less. We have already reported that the molecular C-K edge spectra can be used to predict the molecular properties [2], and we have recently published these organic molecular C-K edge spectral database [3]. We expect that this database will be used as benchmarks for organic molecule spectral informatics or for the analysis of experimentally obtained spectra.<br/>In this presentation, we present the results of developing a ML model using a graph neural network with the organic molecular structures and an excitation carbon site therein as input and the corresponding site C-K edge spectra as output. After training on the database, our ML model is generally successful in reproducing spectral outlines such as rough features and major peaks.<br/><br/>References<br/>[1] R. Ramakrishnan, P. Dral, M. Rupp, and O. A. Lilienfeld, <i>Sci Data</i> <b>1</b>, 140022 (2014).<br/>[2] K. Kikumasa S. Kiyohara K. Shibata, and T. Mizoguchi, <i>Adv. Intell. Syst.</i> <b>4</b>, 2100103 (2021).<br/>[3] K. Shibata, K. Kikumasa, S. Kiyohara, and T. Mizoguchi, <i>Sci. Data</i> <b>9</b>, 214 (2022).