Small Data is All You Need – Machine Learning-based Feature Identification in Scanning Transmission Electron Microscopy (STEM) Images of MoS2

Le résumé fourni par la source

Machine learning (ML) has transformed many fields, but its impact in materials science and microscopy is limited by the scarcity of high-quality labeled data. Unlike domains with abundant datasets, microscopic data availability is constrained by expensive instrumentation, highly specific technical domains, and labor-intensive labeling. This gap poses a challenge for conventional ML models trained on large amount of labeled data to be deployed in electron microscopy. Overcoming these limitations requires targeted small-data machine learning approaches through careful feature engineering and data augmentation, intelligent choice of model and training approaches, and technical understanding of the scientific problem in hand. In this work, we showcase our exploration and achievement in addressing this small-data challenge with a comprehensive machine learning framework for real-time atomic feature identifications of MoS2 monolayer in scanning transmission electron microscopes (STEM). Based on High-Angle Annular Dark Field (HAADF) image of MoS2 acquired real-time on STEM, we establish a comprehensive machine learning framework combining ensemble learning on simulated data [1], robust random forest model trained on limited experimental data, and convolutional neural networks (CNN) to efficiently and accurately identify features of interest. We will also share our explorations on the choice of embedding algorithms for STEM data and its effect on the performance of our final model in accurately labeling atomic features. Finally, we demonstrate the application of our machine learning based feature detection framework by incorporating them into an autonomous atom-by-atom fabrication workflow, where we strategically manipulate localized beam exposure on atoms for controlled growth of Mo6S6 nanowire structure on pristine MoS2 monolayer surface [2]. Visualization of our feature identification workflow, which consists of an atom finding model trained through ensemble learning, convolutional neural network (ConvNet) based nanowire (NW) identification model, and a random forest model to identify regular Mo, S, and Mo single vacancy lines (SVL). Autonomous workflow of targeted MoS-NW growth incorporating our machine learning feature classification framework.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.