Lauri Kurki1, Niko Oinonen1,2, Adam S. Foster1,3
1Aalto University, Espoo, Finland
2Nanolayers Research Computing Ltd., London, UK
3WPI Nano Life Science Institute, Kanazawa, Japan
Scanning tunnelling microscopy (STM) and atomic force microscopy (AFM) functionalized with a CO molecule on the probe apex capture sub-molecular level detail of the electronic and physical structures of a sample from different perspectives [1]. However, the produced images are often difficult to interpret with respect to both physical and chemical structure. To accelerate the analysis, we propose automated machine learning image interpretation tools to extract sample properties directly from bond-resolved STM images [2]. In recent years, there has been rapid development in image analysis methods using machine learning with particular impact in medical imaging. These concepts have been proven effective also in SPM in general and especially for extracting sample properties from AFM images [3,4,5]. We build upon these models utilising convolutional neural networks for image analysis, and show that we can extract atomic positions directly from STM images. Finally, we establish the limits of the approach in an experimental context by predicting atomic structures from STM images of various small organic molecules. We also test the chemical sensitivity of the method by predicting chemical compositions of some organic molecules.
[1] Cai, S., Kurki, L., Xu, C., Foster, A. S., Liljeroth, P., J. Am. Chem. Soc. 144, 20227-20231 (2022).
[2] Kurki, L., Oinonen, N., Foster, A.S., ACS Nano, 18, 17, 11130–11138 (2024).
[3] Alldritt, B., Hapala, P., Oinonen, N., Urtev, F., Krejci, O., Canova, F. F., Kannala, J., Schulz, F., Liljeroth, P., Foster, A. S., Sci. Adv. 6, eaay6913 (2020).
[4] Carracedo-Cosme, J., Perez, R., npj Comput. Mater. 10, 19 (2024).
[5] Oinonen, N., Kurki, L., Ilin, A., Foster, A.S., MRS Bulletin 47, 895-905 (2022).