Scanning probe microscopy (SPM) has been a premier tool in the nanoscientist’s toolbox for the past three decades and combines the ability to detect minute forces with imaging at the atomic scale. As such, it has been heavily used to both probe and modify materials to better understand how functional properties are related to microstructural features of samples.
In the past decade, machine learning (ML) methods have become adopted across the microscopy community, and in the case of SPM, have been used primarily to improve data analysis after the fact. However, one major advantage of recent ML and deep learning approaches is that they can be rapid, and thus be used online during acquisition to improve the data capture process, and to autonomously “drive” the instrument, turning the SPM from a human-operated characterization tool to a “smart” scientific discovery platform capable of autonomous discovery. In this talk, I will discuss the history of ML in SPM, starting with our efforts at the Center for Nanophase Materials Sciences. We will begin with a review of how linear and nonlinear unmixing approaches can be used for unsupervised machine learning to discover physics from spectroscopic datasets, enabling visualization and identification of unique spectroscopic signatures in ferroelectric materials. Next, I will explain the use of modern deep learning and Bayesian optimization approaches to not only improve signal acquisition and analysis, but to be integrated into the microscope for feedback and autonomy. Finally, we will explore the intersection between reinforcement learning, automated and autonomous SPM, and coupled computational-experimental workflows, and suggest that SPM can be a crucial platform for testing and deploying state of the art ML algorithms.
This work was supported by the Center for Nanophase Materials Sciences, which is a US Department of Energy, Office of Science User Faciltiy at Oak Ridge National Laboratory.
Copyright. All Rights Reserved.