Reinforcement Learning for Electron Microscopy Automation

Theory, synthesis, and characterization are essential components of the materials informatics cycle. While significant emphasis has been placed on theory and synthesis, comparatively limited progress has been made in materials characterization. Initial efforts to apply machine learning to characterization have mainly focused on the backend data analysis, but capturing reproducible data in a statistically significant way remains a major challenge. Although electron microscopy can provide atomic-level structural measurements of materials to unravel structure-property relationships, some critical limitations to the current workflow severely limit the current usage and future potential for materials informatics. The largely manual nature of the technique requires a significant amount of time to characterize an extremely small volume of material. This makes it challenging to collect extensive enough datasets that truly represent the material, especially when the material is inhomogenous or contains various defects. Moreover, the human input in collecting images and spectroscopic data inherently has bias and random error that is undesirable for any comprehensive and systematic study.<br/><br/>In this talk, we will highlight the application of reinforcement learning (RL) techniques in automating and optimizing operations within electron microscopy. RL has proven its capabilities in surpassing human performance in complex systems, including high-skill games. In the context of electron microscopy, RL offers a reliable approach to enhance and automate materials characterization through instrument control. While significant progress has been made in theory and synthesis within the materials informatics cycle, materials characterization remains challenging with comparatively fewer advancements. Current workflows suffer from limitations such as manual data collection, time-consuming processes, and inherent biases and errors. We present machine learning-assisted electron microscopy to address these challenges. By creating a virtual RL environment, we develop a network capable of autonomously aligning the electron beam without prior knowledge. Through extensive simulation and experimental validation, we showcase the robustness and effectiveness of our approach, emphasizing the significance of appropriate virtual environments. The results highlight the potential of RL in streamlining electron microscopy workflows, reducing algorithm design complexities, and facilitating the integration of machine learning techniques for accelerated discovery.