Our new paper on the prediction of cation stoichiometry via machine learning analysis of RHEED patterns is out in Nano Letters. This work was co-led by Sumner Harris at Oak Ridge National Lab and our own Patrick Gemperline as part of his Ph.D. thesis at Auburn. Credit also goes to Rama Vasudevan and Chris Rouleau at ORNL for their guidance on design of the experiments and machine learning approaches.
In this paper, Sumner and Patrick used a gated convolutional neural network (CNN) for a quantitative image regression task, predicting the Sr:Ti ratio in SrTiO₃ films. They found that this works well with a small dataset (31 samples) and using a small model (900k parameters), returning predictions within the usual uncertainty of stoichiometry when measured by X-ray photoelectron spectroscopy. In the process, the neural network also demonstrated that the 2nd order diffraction peaks carry key information about the Sr to Ti ratio in the films.
Since RHEED is so central to MBE and PLD synthesis, this is a big step forward for our ultimate goal of enabling autonomous synthesis of films via a ML feedback loop that monitors RHEED and other system parameters. There will be more to come as we expand our data science efforts at UD and continue to collaborate with Oak Ridge and other partners.
This work was funded by the National Science Foundation through our CAREER Award and the Department of Energy through the Center for Nanophase Materials Sciences and Patrick’s Office of Science Graduate Student Research (SCGSR) Fellowship.