Ehsan Khatami, a professor of physics and astronomy in the College of Science, has his latest research published in the journal Nature today in an article entitled “Machine Learning in Electronic-Quantum-Matter Imaging Experiments.” The article shares research that is a collaboration between Khatami and colleagues at Cornell University as well as other institutions on AI-assisted discovery in images of an electronic order in a superconducting material.
A former student of Khatami’s, Kelvin Chng, is cited as an author on the paper and is now working for an AI company in the Bay Area.
“When I first heard about preliminary applications of machine learning methods in condensed matter physics at a conference in spring of 2016, I did not know anything about them,” he said. “I came back from the conference with some ideas on how to use them for quantum problems and quickly found out that Kelvin, who was working on a different project at the time, already knew a lot about artificial intelligence and their applications in industry.”
The pair began working and by the end of the summer had completed a paper that was published a year later in Physical Review X and highlighted in the American Physical Society News. They began collaborating with the Cornell group on designing and testing machine learning algorithms to categorize quantum electronic images of superconducting materials called cuprates.
The images were taken at Cornell using a method called scanning tunnelling microscopy, which maps out real-space patterns of electrons that have self organized into complex quantum states. The images are so noisy and naturally chaotic that conventional methods, like the Fourier analysis, have not been able to decisively pinpoint the type of electronic order found in samples that are close to becoming superconductors in a state dubbed by some physicists the ‘dark matter’ of cuprates.
The team used machine learning for the first time to make sense of data in this mysterious region. They trained a group of artificial neural networks using images that were generated via computer models based on a set of hypotheses, and found that the networks consistently discover the predominant features of a specific ordering pattern whose description dates back to the 1990s.
“It took a long time, many trials, and months of hard work for the collaboration at the beginning to carve out the best strategy to approach and solve the problem, but it all paid off over two years later.”