Autonomous Tip Preparation for STM using Machine Learning on Simulated Data

Abstract

An important time consuming task in scanning tunneling microscopy (STM) is to prepare the tip for experiments. That is why the last few years several research groups have attempted to develop an autonomous tip preparation method using a training database with experimental STM scans. In this thesis, the development of an autonomous tip preparation method using simulated data is described. Muffin- Tin simulations are used to generate a basic simulated STM scan of Cu(111) with CO molecules damped on the surface. This STM scan is manipulated with various fi?lters to create different STM scan categories so that an extensive and consistent training database with simulated data is realized. Open-source machine learning libraries Tensor flow and Keras are introduced using Python 3.8 to train an ensemble of convolutional neural networks on the simulated database. Using an ensemble of 30 networks, we achieve a precision of 94% and a recall of 80% on experimental images, with a threshold of 65%. Initial tests of an autonomous tip-preparation tool based on this ensemble of neural networks provides an atomically sharp tip in 90.6% of the ?nal stops. Further research will focus on further improving the neural network and on expanding the methodology to other substrates.