Automated classification of individual atoms on surfaces using machine learning

TL;DR

This paper presents a machine learning approach using convolutional neural networks to automate the classification of individual atoms on surfaces via STM images and spectroscopy, significantly improving speed and accuracy.

Contribution

The study introduces a semi-automated ML method for classifying surface atoms, achieving high accuracy and demonstrating applicability to various nanoscopic measurements.

Findings

Topography model achieves 86% validation accuracy

Spectroscopy model achieves 100% accuracy

Method applicable to atom and molecule classification on surfaces

Abstract

Leveraging scanning tunneling microscopy (STM) for atomic-scale fabrication has led to many advancements such as the creation of atomic electron-spin qubit structures on surfaces. However, the time-consuming and tedious nature of this process calls for improvements, and this study explores the use of machine learning (ML) to automate certain steps, notably identifying appropriate atomic candidates for the structures. We classify titanium and iron atoms on a magnesium oxide (MgO) surface, which are prototypical on-surface spin qubit candidates, showing distinct topographic and spectroscopic features depending on the bonding sites of the MgO surface. Employing a semi-automated computer vision process, we train a convolutional neural network with STM topographic images and scanning tunneling spectroscopy (STS) curves of several hundred atoms. After training, the topography model achieves an 86% validation accuracy in classifying 200 new images, and the STS model, a 100% accuracy for a sample size of 100 atoms. This method extends its applicability to various nanoscopic measurements, including atomically resolved imaging and local probing of electronic properties, offering a promising approach for classifying atoms and molecules on surfaces.