Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 35: Organic molecules on inorganic substrates: Adsorption and growth II
O 35.7: Vortrag
Dienstag, 10. März 2026, 12:00–12:15, HSZ/0401
Using Machine Learning to predict molecular conformations from STM images — •Tim J. Seifert1, Dhaneesh Kumar2, Stephan Rauschenbach3, Klaus Kern2, Markus Etzkorn1,4, Kelvin Anggara2, and Uta Schlickum1,4 — 1Institute of Applied Physics, TU Braunschweig, Braunschweig — 2Max Planck Institute for Solid State Research, Stuttgart — 3Department of Chemistry, Kavli Institute for Nanoscience Discovery, University of Oxford, Oxford, United Kingdom — 4Laboratory for Emerging Nanometrology LENA, Braunschweig
Single-molecule imaging of biologically essential building blocks such as saccharides and peptides by scanning tunneling microscopy (STM) holds immense promise, enabling key insights into their structure and biological functions. However, automatic structural analysis remains bottlenecked by the lack of available experimental training data. We overcome this fundamental limitation through a novel workflow that rapidly generates vast datasets of high-fidelity synthetic STM images from geometric molecular models. We train a custom Machine Learning architecture on this synthetic data to predict atomic coordinates directly from STM images, enabling fully automated reconstruction of molecular conformations. Validated on two distinct organic systems - glycans and polypeptides - the method achieves atomic positional accuracies below 4 Å and 2 Å, respectively, with unsupervised conformer classification of polypeptides exceeding 97% accuracy. The trained models transfer robustly to experimental STM data, delivering visually convincing structural predictions.
Keywords: Synthetic data generation; Machine Learning structure prediction; Scanning Tunneling Microscopy; Computer Vision