Autonomous Classification of Scanning Tunneling Spectroscopy via Deep Learning with Variational Autoencoders

Scanning tunneling spectroscopy (STS) is one of the primary measurement techniques of scanning tunneling microscopy (STM) which provides information on the electronic structure of the sample. It is one of the few spectroscopic methods which allows the probing of the local density of electronic states above the fermi level. We have demonstrated the autonomous classification of STS of several material systems using deep learning. Specifically, a beta variational autoencoder was trained on un-labled STS data from the reconstructed Au(111) and Si(100) surfaces, as well as pristene and defected tungston disulfide. Importantly, this un-supervised method does not require the user to manually give labels to hundreds or thousands of spectra to train their deep learning model. This avoids label bias and a major throughput bottle neck when the material under study changes and new STS must be labeled. Finally, this method opens up a new avenue of tracking STM tip state changes during experiments.