The Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship

Yang Long

Host School ​: SPMS​
Email: [email protected]
Google Scholar: https://scholar.google.com/citations?user=D_YienkAAAAJ
ORCID: https://orcid.org/0000-0001-7600-3396
Github: https://longyangking.github.io/

Yang Long obtained his Ph.D. in Physics from Tongji University, Shanghai, China in 2020. Currently, he is a research fellow at Nanyang Technological University, Singapore. Dr. Long's research expertise includes investigating spin angular momentum and new topological phenomena in classical wave systems, directional control of near-field photonics/phononics, and machine learning in fundamental physics. He is currently concentrating on the classification of topological phases of matter without human knowledge and developing new kinds of unsupervised learning algorithms for replacing complicated abstract mathematics.

• Machine learning in physics
• Topological physics
• Near-field wave physics
• Spin angular momentum in classical waves

Project: Machine learning of topological classifications without human knowledge
Abstract: A significant and unfading theme in fundamental science is the classification of matter. The classifications of matter reflect the uses and functions of matter. However, the traditional classifications of matter are primarily restricted to using a few common physical characteristics or degrees of freedom, such as elements, geometrical structures, or chemical compounds. Topological classification, a new classification scheme based on topology, has been gradually introduced into investigations of classifications of matter. Natural matter can be further divided into various topological classes, such as topological/trivial insulators and topological semimetals, in accordance with topological classifications. Novel topological phenomena between two topologically distinct materials have been observed with plenty of applications such as topological laser and on-chip THz-region optical transport. Although many theories about topological classifications have been constructed, there are still new topological mechanisms or materials reported very recently. The reason is that these theoretical approaches heavily rely on abstract mathematics (such as homotopy group, K-theory, or Clifford algebra), which is incomplete, under slow development, and may be incorrect in new cases. Note that many materials previously classified as “trivial” are found to be topological with new theoretical advances. Therefore, a question is raised: is it possible to dispense with flawed mathematical techniques (e.g., homotopy group and K-theory) or refrain from using any human knowledge? In other words, can we realize topological classifications only based on raw data collected from matter? In this project, I aim to systematically incorporate machine learning to realize topological classification in a data-driven manner, in order to abandon the traditional abstract mathematics. The methodology is based on applying unsupervised learning technologies to raw data from Hamiltonian samples to capture topologically distinct phases. In contrast to earlier math-heavy theories, this data-driven approach will not require any human experience or knowledge. Therefore, it will not miss a “hidden” topological phase due to the incomplete list of topological invariants or theoretical limitations. It also takes practical constraints (e.g., a finite number of bands) into account, and reveals some previously unnoticed features.