Machine Learning-Driven Automated Scanning Probe Microscopy

Y. Liu, K.P. Kelley, R.K. Vasudevan, M. Ziatdinov, S.V. Kalinin
Oak Ridge National Laboratory,
United States

Keywords: autonomous experiments, machine learning, scanning probe microscopy


Scanning probe microscopy (SPM) has become a mainstay of the fields of materials science, condensed matter physics, and so on. However, until now, the search for interesting functionalities in microscopy experiments has depended on human-based decision making, i.e., operators determine the parameters for subsequent experiments according to the previous experiments, to identify objects of interest and the exploration of physical mechanisms. Machine learning (ML) have rapidly become an indispensable part of physics research to explore the physical mechanisms encoded in microscopy data. Here, we developed ML-driven automated experiment (AE) SPM workflow to learn the functionality and mechanism of materials in an automatic manner during experiments. We implemented three machine learning models—including a deep residual learning framework (Res) combining holistically-nested edge detection (Hed), deep kernel learning (DKL), and hypothesis learning—based workflows in piezoresponse force microscopy (PFM) to investigate ferroelectric materials. First, the ResHed converts the data stream from real-time operating SPM into the semantically segmented image of domain walls.1 With pre-selected experimental workflows on thus discovered domain walls, we observed alternating high- and low- polarization dynamic ferroelastic domain walls in a (PbTiO3) PTO thin film. Second, the DKL actively learns the relationship between structural elements in images and properties encoded in spectra during experiments.2 We observed larger hysteresis opening near 180o domain walls due to the larger polarization mobility in the vicinity of the 180o walls in a PTO sample in DKL-AE. Third, the hypothesis learning method identifies the best physical models that can describe the material behavior in an automated manner during the experiment.3 We used this method to investigate the domain growth mechanism in PFM, and it revealed that the domain growth is ruled by kinetic control in a BaTiO3 thin film. These approaches can also be used to investigate other phenomena, including domain wall dynamics, the conductivity of topological defects, and the relationship between domain structure and local properties. We implemented these approaches in SPM here, however, these approaches be adapted to apply to a broad range of imaging and spectroscopy methods, e.g., electron microscopy, optical microscopy, and chemical imaging, so as to apply to a broad range of physical and chemical microscopy experiments. Acknowledgements: This work (implementation, measurement, and data analysis) was primarily supported by the center for 3D Ferroelectric Microelectronics (3DFeM), an Energy Frontier Research Center funded by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences under Award Number DE-SC0021118. This work (ML development) was supported by the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility. References: [1] Liu, Yongtao, Kyle P. Kelley, Hiroshi Funakubo, Sergei V. Kalinin, and Maxim Ziatdinov. Advanced Science 9, no. 31 (2022): 2203957. [2] Liu, Yongtao, Kyle P. Kelley, Rama K. Vasudevan, Hiroshi Funakubo, Maxim A. Ziatdinov, and Sergei V. Kalinin. Nature Machine Intelligence 4, no. 4 (2022): 341-350. [3] Liu, Yongtao, Anna Morozovska, Eugene Eliseev, Kyle P. Kelley, Rama Vasudevan, Maxim Ziatdinov, and Sergei V. Kalinin. arXiv preprint arXiv:2202.01089 (2022).