R. Vasudevan, B. Smith, A. Khojandi, S.M. Valleti, S. Jesse, Y. Liu, S.V. Kalinin, M. Ziatdinov
Oak Ridge National Laboratory,
United States
Keywords: machine learning, piezoresponse force microscopy
Summary:
The combination of automation of microscopy platforms, in combination with robust and trainable machine learning models, has enabled autonomous microscopy to become feasible in the last few years. These autonomous systems leverage the advantages of decision-making algorithms such as reinforcement learning or Bayesian optimization methods to deal with sequential decision-making tasks in uncertain environments to optimize properties of interest. In this talk, we will review recent advances in the autonomous microscopy platforms at the Center for Nanophase Materials Sciences, focusing on the use of such systems to characterize and manipulate dynamics in thin film ferroelectric materials. Specifically, we will begin with how automated and autonomous spectroscopies can be used to explore structure-property relationships, in real time, on the microscope in an efficient manner, utilizing less than 20% of the total time for a usual grid-based method. In the process, the local descriptors that characterize dynamic behaviors are learned via neural nets as part of a deep kernel learning framework. Next, we will explore how automation of data collection can enable building of surrogate models that can describe system dynamics for different starting configurations and bias pulse parameters. These surrogate models are then used within reinforcement learning environments, that can be used to train reinforcement learning agents that can learn policies to optimize domain structures based on targeted structures and/or functionalities. Extensions to curiosity-driven RL frameworks implemented on the microscope are discussed. This shows the potential to use autonomous SPM approaches not just to better understand materials systems, but to actively manipulate them towards desired structures, which would be difficult or impossible to achieve with standard human operation. This work was supported by the Center for Nanophase Materials Sciences, which is a US Department of Energy, Office of Science User Facility at Oak Ridge National Laboratory.