Scanning tunneling spectroscopy (STS), based on scanning tunneling microscopy (STM), is a powerful tool to characterize the electronic structure of single molecules and nanomaterials. While performing STS, the structure and condition of the probe tips are critical for obtaining reliable and stable spectra. Here, we present an automated program based on machine learning models that can identify the Au(111) Shockley surface state from dI/dV spectra and perform tip conditioning on clean or sparsely covered gold surfaces with little user intervention. We employ a straightforward height-based segmentation algorithm to analyze STM topographic images and identify tip conditioning positions and used 1789 archived dI/dV spectra to train machine learning models that can predict the condition of the tip by evaluating the quality of the spectroscopy data. Decision tree based ensemble and boosting models and deep neural networks (DNNs) have been proven to perform reasonably on identifying tips in usable conditions for STS. A generalizable data augmentation algorithm has been developed to generate artificial dI/dV spectra using fewer than 100 spectra, which are then used to train DNNs that can identify better dI/dV spectra for gold. However, overfitting problems exist in these models for artificial spectra and advanced machine learning techniques are required for further improvements. We expect the program to save human labor, reduce research costs for surface science studies and accelerate the discovery of novel nanomaterials by STM. The strategies we used can also be generalized for STM tip conditioning on other metallic surfaces.

Automated tip conditioning for scanning tunneling spectroscopy

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.