Machine learning enhanced scanning tunneling microscopy (STM)

Abstract

Atomic-scale surface characterization is critical in current scientific research, with scanning tunneling microscopy (STM) providing high-resolution topography images and spectroscopic data that reveal the quantum behavior of materials. However, STM's effectiveness depends heavily on the quality of its atomically sharp tip, which is prone to contamination and damage from atoms, molecules, or surface debris, frequently demanding manual human calibration. This project aims to increase the operational efficiency and automation of the STM by combining machine learning (ML) and artificial intelligence (AI) to allow the autonomous calibration of the tip, which will serve as a proof of concept for automated ML-enhanced STM operation in our laboratory. The study investigates two methods using a compiled dataset of a known crystalline surface, Au(111), recognized for its unique herringbone structures. Method 1 utilizes the DeepSPM framework, obtaining 76.8% accuracy and an AUC of 0.871 in image classification. However, it demonstrated shortcomings in real-time tip conditioning. Method 2, a simpler custom ML approach that employs binary cross-entropy (BCE) loss - BCEWithLogitsLoss, achieved a greater accuracy of 89.45% , and successfully implemented an autonomous tip calibration technique. Despite limitations such as a small dataset and real-time conditioning limits in methods adopted, this project demonstrates the feasibility of applying ML and AI to STM tip calibration. As the first practical application of these AI approaches within our Nanyang Technological University (NTU) STM laboratory, it lays the groundwork for future advancements in fully automated nanoscale imaging and operational efficiency.