The machine learning method of exploiting large volumes of X-ray data will accelerate the discovery of materials.
Color coding makes aerial maps much more understandable. Through color, we can tell at a glance where there is a road, forest, desert, city, river or lake.
Working with several universities, the US Department of Energy’s (DOE) Argonne National Laboratory has devised a method for creating color-coded graphs of large volumes of data from X-ray analysis. This new tool uses sorting computational data to find clusters related to physical properties, such as an atomic distortion in a crystal structure. It should greatly speed up future research into atomic-scale structural changes caused by different temperatures.
“Our method uses machine learning to rapidly analyze large amounts of X-ray diffraction data,” said Raymond Osborn, senior physicist in Argonne’s Materials Science division. “What might have taken us months in the past now takes about a quarter of an hour, with much finer results.”
For more than a century, X-ray diffraction, or XRD, has been one of the most fruitful of all scientific methods for analyzing materials. It has provided key information on the 3D atomic structure of countless technologically important materials.
In recent decades, the amount of data produced in XRD experiments has increased dramatically at large facilities such as the Advanced Photon Source (APS), a DOE Office of Science user facility at Argonne. However, analysis methods that can handle these infinite data sets are sorely lacking.
The team calls their new method X-ray Temperature Clustering, or XTEC for short. It accelerates material discoveries by rapidly collecting and color-coding large X-ray data sets to reveal previously hidden structural changes that occur when temperature is increased or decreased. A typical large data set would be 10,000 gigabytes, equivalent to approximately 3 million songs of streaming music.
XTEC is based on the power of unsupervised machine learning, using methods developed for this project at Cornell University. This machine learning does not depend on initial training and learning with already well studied data. Instead, it learns by finding patterns and clusters in large data sets without such training. These patterns are then represented by color coding.
“For example, XTEC can assign red to dataset one, which is associated with a certain property that changes with temperature in a particular way,” Osborn said. “Then group two would be blue and connected to another property with a different temperature dependence, and so on. The colors indicate whether each group represents the equivalent of a road, forest, or lake on an aerial map.”
As a test case, XTEC analyzed data from beamline 6-ID-D at APS, obtained from two crystalline materials that are superconducting at temperatures near absolute zero. At this extremely low temperature, these materials go into a superconducting state, offering no resistance to electric current. More important to this study, other unusual features appear at higher temperatures related to changes in the material’s structure.
By applying XTEC, the team extracted an unprecedented amount of information about changes in atomic structure at different temperatures. They include not only distortions in the regular arrangement of atoms in the material, but also the fluctuations that occur when such changes occur.
“Because of machine learning, we are able to see the behavior of materials invisible from conventional XRD,” Osborn said. “And our method is applicable to many big data problems not only in superconductors, but also in batteries, solar cells and any temperature-sensitive device.”
APS is undergoing a massive upgrade that will increase the brightness of its X-ray beam by up to 500 times. Along with the upgrade will come a significant increase in the data collected in APS, and machine learning techniques will be essential for analyzing this data in a timely manner.
In addition to Osborn, Argonne authors include Matthew Krogstad, Daniel Phelan, Puspa Upreti, Michael Norman, and Stephan Rosenkranz. The main collaborating partner is Cornell University (Eun-Ah Kim, Jordan Venderley, Krishnanand Mallayya, Michael Matty, Geoff Pleiss, Varsha Kishore and Kilian Weinberger) and the Cornell High Energy Synchrotron Facility (Jacob Ruff). Other partners include the University of Tennessee (David Mandrus), the University of Maryland (Lekh Poudel) and New York University (Andrew Gordon Wilson).
Argonne funding was provided by the DOE Office of Basic Energy Sciences and the National Science Foundation.