Presentation Information
[1P09]Automated detection of topological features in fermi surfaces via interpretable machine learning
*Daichi Ishikawa1, Kentaro Fuku1,2, Yoshio Miura3,4, Yasuhiko Igarashi5, Yuma Iwasaki6, Yuya Sakuraba3, Koichiro Yaji6,7, Alexandre Lira Foggiatto1, Takahiro Yamazaki1, Naoka Nagamura1,6,8, Masato Kotsugi1 (1. Department of Material Science and Technology, Tokyo University of Science, 2. Department of Chemistry, Graduate School of Science, Nagoya University, 3. Research Center for Magnetic and Spintronics Materials (CMSM), National Institute for Materials Science (NIMS), 4. Electrical Engineering and Electronics, Kyoto Institute of Technology, 5. Institute of System and Information Engineering, University of Tsukuba, 6. Center for Basic Research on Materials (CBRM), National Institute for Materials Science (NIMS), 7. Unprecedented-scale Data Analytics Center (UDAC), Tohoku University, 8. Research Institute of Electrical Communication (RIEC), Tohoku University)
We applied unsupervised machine learning to the Fermi surface of Co2MnGaxGe1-x (CMGG) and graphene on SiC to automatically extract physical property–related features. Band structures of CMGG were calculated using VASP and projected onto the kx-ky plane, and spin polarization was evaluated from the density of states. Principal component analysis (PCA) reduced the high-dimensional Fermi surface data, visualizing composition-dependent changes and detecting significant “jumps” corresponding to extrema in spin polarization or the appearance of gapped nodal lines. PCA also revealed deviations from data trends at specific compositions and identified Dirac points in graphene ARPES data. Robustness tests confirmed that key features remained detectable under added noise and broadening. These results demonstrate that unsupervised learning effectively captures subtle Fermi surface variations and is promising for analyzing noisy ARPES data to uncover hidden electronic structures.