We’ve been working in earnest for several years on machine learning and data analytics for maximizing the information we glean from reflection high-energy electron diffraction (RHEED). In the MBE and PLD world, RHEED is used to monitor the growth of epitaxial films in real time, generating information on growth rates and surface quality. If you spend enough time staring at a RHEED screen though, you quickly learn that there is a lot more information there that is very difficult to quantify and very difficult to interpret, but still empirically valuable to the film grower. This sets up really well for machine learning if you have enough data and good enough data to train models with. For your consideration, we present some new tools for the thin film community in our new paper out in Journal of Vacuum Science & Technology A today as an Editor’s Pick. Patrick Gemperline led this work as part of his Ph.D. thesis at Auburn in collaboration with Rama Vasudevan at ORNL and working off of RHEED videos from LaFeO3 films grown by Rajendra Paudel during his time at Auburn.
Through a new algorithm for RHEED image alignment which we call residual sum of squares (RSS), we can now overlay videos from multiple growths and process them together with Python analytics codes for principal component analysis and K-means clustering. That means if you want to do a long-term comparison across a series of samples (i.e. different doping levels, different growth temperatures, on vs. off-stoichiometry, etc), you can do that. We have also demonstrated ways to non-linearly tune intensity that emphasize subtle features such as Kikuchi bands. The codes are freely available on our Github.
This work was supported by the National Science Foundation (NSF) for our group and U.S. Department of Energy Office of Science for Patrick’s fellowship to work at Oak Ridge National Laboratory in the Center for Nanophase Materials Science.