Machine learning for many-Body physics

ORAL

Abstract

We investigate the application to many-body physics of Machine Learning (ML) methods for predicting new results from accumulated knowledge. We show that ML can be used efficiently for the Anderson impurity model (AIM)[1] and present preliminary results on its use as a solver for dynamical mean field theory (DMFT). We establish that the best representation of the Green's function for ML is by parametrizing it as an expansion in term of Legendre polynomials [1]. In DMFT applications, a key issue is the choice of descriptor, the data representation used as input for ML, which is not dependent on the impurity solver. Different parametrizations are examined. The ability to distinguish metallic and Mott insulating solutions is analysed.\\[4pt] [1] L.-F. Arsenault et al., PRB 90, 155136 (2014)

*DOE No. 3F-3138

–

Authors

  • Louis-Francois Arsenault

    • Department of Physics, Columbia University, New York, New York 10027, USA

  • Alejandro Lopez Bezanilla

    • Materials Science Division, Argonne National Laboratory, 9700 S. Cass Avenue, Lemont, IL 60439, USA
    • Argonne Natl Lab
    • Argonne National Lab

  • O. Anatole von Lilienfeld

    • Institute of Physical Chemistry, Department of Chemistry, University of Basel, Klingelbergstrasse 80, CH-4056 Basel, Switzerland

  • Andrew Millis

    • Columbia University
    • Department of Physics, Columbia University, New York, New York 10027, USA
    • Departemtn of Physics, Columbia University
    • Dept. of Physics, Columbia Univ.