Automated image analysis for single molecule manipulations via STM
Michael Obermayr
https://tugraz.webex.com/tugraz/j.php?MTID=m2790c3d2a9a49c31369978e2a76de31c
11:15 - 12:15 Wednesday 10 April 2024 Scanning tunneling microscopy (STM) allows precise manipulation of single atoms and molecules on surfaces. Recent AI-driven advancements in manipulating arbitrary molecules [1] open the door to automatic assembly of artificial nanostructures. A significant time portion is required for the imaging before and after each manipulation step, which often acts as the speed bottle neck. In this work we train a neural network on the tunneling current to predict the shift in position and orientation of an object, limiting the need for constant imaging.
The challenge in training a neural network lies in obtaining a suitable dataset containing all necessary inputs and outputs. Here we start with a measurement series of single-molecule manipulations paired with the respective precedent and subsequent images of the surface. The action outcome is extracted from these images with machine vision. Thus, a suitable training dataset for neural networks can be generated systematically, while ensuring broad applicability towards arbitrary molecules.
While the AI-driven predictions fall short of completely replacing imaging steps, the resulting dataset allows for a detailed analysis of single-molecule motion on surfaces, unveiling the inherent stochastic nature of the process. The relationships between rotational and translational outcomes during manipulation steps can be studied, in order to understand the underlying mechanisms at play.
[1] B. Ramsauer et al., J. Phys. Chem. A, 127, 2041 (2023)
|
Institute of Solid State Physics
