Machine learning in condensed matter physics

Eliska Greplova, Mark H Fischer and Sebastian Huber
Institute for Theoretical Physics, ETH Zurich, Zurich, Switzerland

Machine learning has been at the forefront of computer science for decades [1] and found an impressive number of technological applications. Artificial intelligence tools in physics (see e.g. [2-4]) have recently been in the centre of scientific interest. Our current goal is to expand the field of machine learning in physics and search for new insights into recent and long-standing problems. We present overview of our current projects in this direction that involve (a) efficient numerical analysis of experimental data, and (b) search for new phenomena in condensed matter physics.

The former employs machine learning based image recognition techniques that serve both for pre-processing of microscope images (such as samples with graphene flakes) in order to identify suitable experimental samples as well as analysis of experimental signals even under the conditions that are too dire for any standard techniques to work [5].
The latter is focused on identifying emergent phenomena in condensed matter physics without any human input.

[1] Rumelhart, DE, Hinton, GE and Williams, RJ, Nature 323, 533 (1986)
[2] Carleo, G and Troyer, M, Science 355, 602 (2017)
[3] van Nieuwenburg, E, Li, Y and Huber, SD Nature Phys. 13 435 (2016)
[4] Koch-Janusz, M and Ringel Z Nature Phys. (2018)
[5] Greplova, E, Andersen, CK and Molmer, K, arXiv:1711.05238 (2017)