Autonomous Elucidation of Structure-Property Relationships in Ferroelectrics via Humanin the Loop Bayesian Optimization, Reinforcement Learning and Scanning Probe Microscopy

Recent progress in the areas of autonomous experiments, particularly at major user facilities, has ushered in a new paradigm over traditional human-based operations, thereby enabling new experiments that were hitherto impossible via traditional operations. To date, however, most autonomous systems in materials characterization and synthesis facilities rely on straightforward Bayesian optimization where targets are pre-defined by the user, and no subsequent interactions between the operator and the instrument are expected.<br/><br/>Here, we will discuss autonomous experiments performed on scanning probe microscopes (SPM) at the Center for Nanophase Materials Sciences. We first show that using human in the loop models, where the operator is presented with a subset of the experimental results that can be voted on, is an effective strategy for leveraging the best features of human-based curiosity along with machine learning -based optimization. This is implemented in the form of a Bayesian optimization active recommender system model, on a functioning scanning probe microscope to investigate relationships between the local domains structure and the characteristic features of ferroelectric hysteresis.<br/><br/>Next, we explore the use of automated and autonomous SPM platforms for (a) developing surrogate models of domain wall dynamics, and (b) training and deploying reinforcement learning (RL) agents to automatically manipulate domain wall structures towards desired morphologies. Of note, these surrogate models are validated via phase-field simulations and reveal peculiar features including effects of local wall stresses on functional responses. The RL agents are implemented on the working instrument, and challenges and future directions of this strategy towards autonomous manipulation of materials is discussed. Overall, this work shows that utility of automated and autonomous SPM platforms, that enable qualitatively new types of experiments and additional physical insights that are difficult to capture through non-autonomous means. This work was supported by Center for Nanophase Materials Sciences (CNMS), which is a US Department of Energy, Office of Science User Facility at Oak Ridge National Laboratory.