2023 Annual Meeting of the APS Mid-Atlantic Section
Friday–Sunday, November 3–5, 2023;
University of Delaware, Newark, Delaware
Session H02: Symposium on 2D and Quantum Materials
11:00 AM–12:48 PM,
Sunday, November 5, 2023
University of Delaware
Room: Gore 104
Chair: Yue Li, Argonne National Laboratory
Abstract: H02.00003 : Bridging Theory with Experiment: Digital Twins and Deep Learning Segmentation of Defects in Monolayer MX2 Phases
12:12 PM–12:48 PM
Abstract
Presenter:
Addis Fuhr
(Oak Ridge National Laboratory)
Author:
Addis Fuhr
(Oak Ridge National Laboratory)
Monolayer MX2 materials (where M indicates a transition metal and X represents chalcogen or oxygen atoms) can exhibit a broad array of physical properties such as metallic, semiconducting, insulating, correlated, and topological phases. In many cases these properties can be considerably altered by the introduction of point defects. However, it is difficult to predict synthesis-structure-property relationships a priori, which limits the prospects of defect engineering nanostructures for various applications such as quantum computing, catalysis, or semiconductor electro-optical devices. High-throughput identification of defects under various synthesis conditions is a principle limitation to determining defect-induced changes in MX2 physics. Major strides have been made by integrating machine learning (ML) methods with scanning transmission electron microscopy (STEM), but segmentation models generally do not transfer well between materials. Moreover, MX2 phases are beam insentitive, which makes it difficult to determine if defects were created during the original synthesis or are beam-induced. Our work highlights the benefits of employing first principles calculations to produce digital twins for training deep learning segmentation models for STEM experiments. We optimized ~600 dffferent defective monolayer MX2 phases with DFT, and simulated their STEM images with multislice simulations for the purpose of generating digital twins. Several deep learning segmentation architectures were trained on this dataset, and their performances evaluated under a variety of conditions such as recognizing defects in the presence of unidentified impurities, beam damage, grain boundaries, and with reduced image quality from low electron doses. This digital twin approach allows benchmarking different deep learning architectures on a theory dataset and the development of new methods for improving segmentation of defects in experiments by enabling the study of defect classification under a broad array of finely controlled conditions. It thus opens the door to resolving the underpinning physical reasons for model shortcomings, methodology to improve segmentation, and potentially chart paths forward for automated discovery of materials defect phases in experiments.