Automated Microscopy for Physics Discovery—From High-Throughput to Hypothesis Learning-Driven Experimentation

In this work, we explore the ferroelectric polarization switching in relation to the applied pulse bias (i.e., bias voltage and time) in an automated manner in scanning probe microscopy (SPM). We perform (1) high-throughput experimentation for a comprehensive understanding of the relationship between pulse biases and ferroelectric domain growth, (2) autonomous experimentation driven by machine learning (ML) algorithm to optimize experimental conditions based on real-time experiment results.<br/><br/>SPM has been a powerful tool for manipulating and visualizing ferroelectric domains at the nanoscale. Investigations of ferroelectric domain size and stability can advance understandings of ferroelectrics application in memory devices, such as operating time, retention time, bit size, and so on. However, these SPM measurements have traditionally been time-intensive and dependent on experienced researchers for monotonous operations and real-time decision-making of measurement parameters, e.g, researchers determine and manually tune the parameters for next iteration of experiment according to the previous results. Here, we perform automated experiments in SPM to explore the mechanism of ferroelectric polarization. First is a high-throughput experiment of writing ferroelectric domains by applying various bias pulse conditions and subsequently conducting piezoresponse force microcopy to image the written domain structure. The high-throughput experimentation enables systematically varying the bias pulse parameters, therefore comprehensively understanding the relationship between the bias parameters and the resulting domain structures. In this experiment, we discovered different polarization states that show up upon different bias conditions. Second, we implement a hypothesis active learning (HypoAL) algorithm to control the SPM for ferroelectric domain writing. The HypoAL is based on structured Gaussian process (sGP), which analyzes the relationship between the bias pulse parameters and the written domain size during operating experiments, and subsequently determines the bias pulse parameters for the next experiment. The HypoAL aims to establish the best physical model from a hypothesis list for the material’s behaviour within the smallest number of experiment step. The HypoAL indicates that the domain growth in a BaTiO₃ film is ruled by kinetic control. We anticipate the established approaches here can be extended to microscopy methods other than SPM in the future to accelerate materials and physics discovery.<br/><br/>Acknowledgement: This work (high-throughput experimentation) is supported by the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility. This work (HypoAL experimentation) is partially supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences Energy Frontier Research Centers program under Award Number DE-SC0021118.