Darian Smalley1,Stephanie Lough1,Luke Holtzman2,Madisen Holbrook2,James Hone2,Katayun Barmak2,Masahiro Ishigami1
University of Central Florida1,Columbia University2
Darian Smalley1,Stephanie Lough1,Luke Holtzman2,Madisen Holbrook2,James Hone2,Katayun Barmak2,Masahiro Ishigami1
University of Central Florida1,Columbia University2
The electronic and optical properties of transition metal dichalcogenides (TMDs) are influenced by the presence of atomic point defects within the crystal lattice. Identifying the quantity and nature of these point defects is important for engineering TMD synthesis, ensuring TMD device consistency, and verifying single photon emission sources. Scanning tunneling microscopy (STM) can determine the nature of such point defects by correlating their electronic structure with atomic-scale spatial properties, which is typically done manually. However, manual identification, counting and measurement of point defects in STM images is painstakingly slow and unreliable, causing reported point defect densities to have large experimental uncertainty, preventing directed refinements of synthesis and controlled reduction in TMD point defect density. We address this problem by processing our STM data with a machine learning algorithm and show atomic-scale defect metrology in scanning tunneling microscopy images of tungsten diselenide using an ensemble of U-Net-like convolutional neural networks. Using this technique, we determined the coordinates, densities, and real space properties of over 3000 defects with statistical significance. Our results demonstrate that STM data analysis aided by machine learning can be used to rapidly determine the quality of TMDs and provide much needed quantitative input to systematically improve the synthesis process. As such our machine learning enhanced STM can impact many two-dimensional materials researchers.