Scanning tunneling state recognition with multi-class neural network ensembles

O Gordon,P Moriarty,L Knijff,F Junqueira,S E Freeney,I Swart,P D’Hondt

doi:10.1063/1.5099590

O Gordon, P Moriarty + Show 5 more

Open Access

https://doi.org/10.1063/1.5099590

Copy DOI

Abstract

One of the largest obstacles facing scanning probe microscopy is the constant need to correct flaws in the scanning probe in situ. This is currently a manual, time-consuming process that would benefit greatly from automation. Here, we introduce a convolutional neural network protocol that enables automated recognition of a variety of desirable and undesirable scanning tunneling tip states on both metal and nonmetal surfaces. By combining the best performing models into majority voting ensembles, we find that the desirable states of H:Si(100) can be distinguished with a mean precision of 0.89 and an average receiver-operator-characteristic curve area of 0.95. More generally, high and low-quality tips can be distinguished with a mean precision of 0.96 and near perfect area-under-curve of 0.98. With trivial modifications, we also successfully automatically identify undesirable, non-surface-specific states on surfaces of Au(111) and Cu(111). In these cases, we find mean precisions of 0.95 and 0.75 and area-under-curves of 0.98 and 0.94, respectively. Provided that training data are available, these ensembles therefore enable fully autonomous scanning tunneling state recognition for a wide range of typical scanning conditions.

Highlights

While scanning tunneling microscopy (STM) has allowed researchers to make observations at the atomic level for decades,1–3 success is highly reliant on the production of atomically sharp scanning tips
All networks performed significantly better than Random Forest Classifier (RFC) and weighted random guessing, the Rashidi and Wolkow8 (RW) Convolutional neural networks (CNNs) performed poorly and similar to the more traditional RFC
At the 32 × 32 image size described by Rashidi and Wolkow,8 RW performed comparable to random guessing, indicating the high difficulty of this task

Summary

Introduction

While scanning tunneling microscopy (STM) has allowed researchers to make observations at the atomic level for decades, success is highly reliant on the production of atomically sharp scanning tips. Time-consuming nature of tip correction, there have been surprisingly few attempts to date to automate the process and allow for the setup and collection of large amounts of data in the absence of a microscopist. Of these attempts, a variety of pitfalls have been identified, ranging from low accuracy and high computational cost to faltering when multiple tip flaws are present. A variety of pitfalls have been identified, ranging from low accuracy and high computational cost to faltering when multiple tip flaws are present They often require a degree of manual input, are invariant to scale and rotation, or fail when the tip spontaneously changes the visible resolution midimage. Convolutional neural networks (CNNs) are highly promising candidates for this task which routinely achieve high accuracy in complex vision tasks such as medical, satellite, and digit recognition. Despite this, in the context of STM, only Rashidi and Wolkow have to the best of our knowledge used CNNs for tip-conditioning and only while scanning the H:Si(100) surface

Methods

Results

Conclusion