Machine learning and artificial intelligence are rapidly becoming an integral part of physics research, with a wide range of applications including theory, materials prediction, and high-throughput data analysis. In recent years, there has been growing interest in using AI models that actively interact with physical systems, rather than being pre-trained on static datasets, for tasks such as materials discovery and optimization, chemical synthesis, and physical measurements. Microscopy is particularly well-suited for these active learning tasks, as it combines aspects of materials discovery through observation and spectroscopy, physical learning with relatively simple prior knowledge and few exogenous variables, and synthesis through controlled interventions. In this presentation, I will discuss recent advances in the use of machine learning for automated experiments in electron microscopy, including object detection, atomic fabrication with electron beams, and physics discovery through active learning. The use of classical deep learning methods in streaming image analysis can be affected by out-of-distribution drift, and I will discuss approaches to circumvent this issue. I will then demonstrate how simple Gaussian processes-based approaches can be suboptimal for active learning in complex systems due to the lack of prior knowledge and overly simplified kernel functions, and how deep kernel learning and structured Gaussian processes can be used to explore complex systems and discover structure-property relationships while enabling automated experiments targeting physics discovery. Finally, I will discuss the high-performance computing and edge infrastructure needs for turning cutting-edge modern-day microscopes into autonomous platforms for scientific discovery.
Back to Learning and Emergence in Molecular Systems
Copyright. All Rights Reserved.