You are here

Machine learning opens new avenues in condensed matter physics research

Monday, February 13, 2017

iStock image - ribbons of binary converging in a pool of lightTwo physicists from the University of Waterloo and Perimeter Institute for Theoretical Physics have succeeded in using machine learning to “recognize” phases of matter. Not only did their neural network exceed expectations, the researchers showed it was able to perform tasks far beyond its original design.

Dr. Roger MelkoWe used one of the simplest machine learning tools and we couldn’t believe how easily this software could be repurposed for our research. We see a lot of potential for machine learning to solve problems in modern condensed matter, where conventional simulation techniques face challenges.

Their proof of concept paper was published this week in Nature Physics.

Roger Melko and his colleague Juan Carrasquilla, a postdoctoral fellow with Perimeter Institute, adapted the Google’s TensorFlow, an open-source neural network, and then fed the software data about a simulated material and its various phases during what’s called a “supervised learning period.”

Supervised learning with neural networks is a cornerstone of machine learning. Take for example the most common form of machine learning today: facial recognition. During the supervised learning step, you tag yourself in a variety of photos. The more snapshots you show the computer, the better it gets at recognizing you in a variety of poses, settings and lighting.

But not only did Melko and Carrasquilla’s neural network perform well at identifying and describing the test material’s phase “snapshots,” the network was able to detect the material’s phase transitions, a task it was never taught.

Among other applications, it means such software could be capable of mapping out phase diagrams of new materials based on limited information, right as they are being created in the lab.

The neural network does have its limitations, however. The system failed when it came to working with complicated quantum wave functions and topological phases – an important avenue of research and the subject of this year’s Nobel Prize.

We’ve adopted a technology from the AI industry with very little of our own modifications,” says Melko. “I want to open the black box and understand why this simple machine learning network works for some phases yet fails to recognise others.

Roger Melko is a professor in the Department of Physics and Astronomy and an Associate Faculty member at Perimeter Institute.

  1. 2018 (7)
    1. March (4)
    2. February (1)
    3. January (2)
  2. 2017 (19)
    1. December (2)
    2. November (3)
    3. October (1)
    4. September (2)
    5. May (1)
    6. April (4)
    7. March (1)
    8. February (3)
    9. January (2)
  3. 2016 (22)
    1. December (2)
    2. November (1)
    3. October (2)
    4. September (4)
    5. July (2)
    6. June (1)
    7. May (1)
    8. April (2)
    9. March (1)
    10. February (5)
    11. January (1)
  4. 2015 (30)
  5. 2014 (12)
  6. 2013 (7)
  7. 2012 (9)
  8. 2011 (3)
  9. 2010 (7)
  10. 2009 (1)