The past decade has seen enormous increases in the size and quality of datasets produced by techniques such as scanning probe microscopy and x-ray diffraction from synchrotrons globally. However, the necessary pathways to both mine the large datasets to derive understanding of fundamental mechanisms, as well as synthesize and compare the results across the wider available literature, are still generally limited. Here, I will present case studies involving our use of machine learning and deep data analysis of scanning probe microscopy datasets for understanding of physical mechanisms.
I will first outline the new suite of techniques that we have developed using scanning probe microscopy, which we term the “General mode” acquisition technique, that streams all of the information available from the measurement system (photodetector, current amplifier, etc.) to be captured and analyzed. This large increase in data volume allows for a wide variety of subsequent analysis, including data-driven filtering, digital lock-ins, and denoising. These techniques can greatly increase the acquisition speed by orders of magnitude for typical SPM experiments, such as in I-V curve acquisition on conductive oxides, and hysteresis loop acquisition for ferroelectrics.
Throughout the talk I will emphasize techniques to learn physics from the large datasets captured, including examples of endmember extraction, Bayesian inference, matrix factorization and convolutional neural networks that can be utilized to automatically learn appropriate features from images for classification, and from which subsequent physics is then derived by combining the information with first principles and thermodynamic models. These advances point to the big-data driven future of scanning probe microscopy as a vital tool for materials science researchers, as a means towards understanding local physics in complex material systems.
This work has been performed in collaboration with Suhas Somnath, Petro Maksymovych, Maxim Ziatdinov, Stephen Jesse, and Sergei V. Kalinin. The imaging and deep data analysis portion was sponsored by the Division of Materials Sciences and Engineering, BES, DOE. This research was conducted and partially supported at the Center for Nanophase Materials Sciences, which is a US DOE Office of Science User Facility.