Physics-Informed Machine Learning Workflows Connecting Electron Microscope to Atomistic Simulations

In recent years, artificial intelligence, and machine learning (AI/ML) methods are being rapidly adapted in physical sciences to gain comprehensive understandings of material structures, properties, system evolutions over spatial-temporal resolution, processes involving phase transitions across various time- and length scales. With the emergence of efficient algorithms and advancement in electron microscopes, there is a scope to utilize theoretical models to guide, perform experiments while refining the parameters in both spaces, to establish a continuous feedback-loop. Instrument specificity, implementation complexity, information transferability by addressing fundamentally different latencies of imaging and simulations, remain the primary challenges. This presentation will focus on how deep learning (DL) frameworks are employed to extract atomic level information from 2D systems such as graphene followed by first-principles based studies to develop comprehensive understanding of the materials physics. These enable seamless deployment of several DL algorithms on-the-fly for appropriate feature finding, property predictions in combination with atomistic simulations to explore underpinning causal mechanisms, suited for guiding next set of experiments. A discussion on extending such workflows for other functional materials with targeted properties on-the-fly will also be included.<br/><br/>This effort (machine learning) is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of Basic Energy Sciences Data, Artificial Intelligence and Machine Learning at DOE Scientific User Facilities, INTERSECT Initiative as part of the LDRD Program of Oak Ridge National Laboratory.