ICN2 Manuel Cardona Lecture: Bits to Atoms and Atoms to Bits (Prof. Sergei V. Kalinin)

Speaker: Prof. Sergei V. Kalinin University of Tennessee, Knoxville, USA
Title: Bits to Atoms and Atoms to Bits: Automated Experiment and Atomic Fabrication in Electron Microscopy
Date: Wednesday 08 March, 2023
Hosted by: ICREA Prof. Jordi Arbiol, Advanced Electron Nanoscopy Group Leader, ICN2.
The Manuel Cardona lecture series offers the ICN2 community and beyond the opportunity to hear from some of the world's top researchers in nano-related fields. They are named in honour of Prof. Manuel Cardona, a Barcelona-born physicist and one of Spain's most prominent scientists who, from his positions abroad, had a profound impact on many of the research centres in our area.
Learn more about the prestigeous ICN2 Manuel Cardona Lectures here: icn2.cat/en/ou...
Abstract: The last note left by Richard Feynman stated “What I cannot create, I do not understand.” Building solid state quantum computers, creating nanorobots, and designing new classes of biological molecules and catalysts alike requires the capability to manipulate and assemble matter atom by atom, probe the resulting structures, and connecting them to macroscopic world. In this presentation, I will discuss recent progress in automated experiment in electron microscopy, ranging from feature and physics discovery via active learning to direct atomic fabrication. I introduce the concept of the reward-driven experimental workflow planning and discuss how these workflows can be implemented via domain-specific hyper languages. The applications of classical deep learning methods in streaming image analysis are strongly affected by the out of distribution drift effects, and the approaches to minimize though are discussed. The real-time image analysis allows spectroscopic experiments at the predefined features of interest and atomic manipulation and modification with preset policies. I further illustrate ML methods for autonomous discovery, where the microstructural elements maximizing physical response of interest are discovered. These deep kernel learning (DKL) methods offer significant advantage compared to simple Gaussian Processes often tend to produce sub-optimal results due to the lack of prior knowledge and very simplified (via learned kernel function) representation of spatial complexity of the system. The DKL AE is illustrated via experimental discovery of the edge plasmons in STEM/EELS. The forensic analysis of the automated experiment makes the discovery process explainable and allows for human in the loop interventions. Finally, I will discuss the opportunities and strategies for direct atomic fabrication via electron beams, targeting desired structures and desired functionalities.