Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 28: Organic Molecules on Inorganic Substrates I: Switching and Manipulation
O 28.10: Talk
Tuesday, April 2, 2019, 12:45–13:00, H24
Machine learning for single molecule manipulation — Philipp Leinen1,2, Malte Esders3, Kristof Schütt3, Klaus-Robert Müller3, F. Stefan Tautz1,2, and •Christian Wagner1,2 — 1Peter Grünberg Institut (PGI-3), Forschungszentrum Jülich, Germany — 2JARA Fundamentals of Future Information Technology, Jülich, Germany — 3Institut für Softwaretechnik und Theoretische Informatik, Technische Universität Berlin, Germany
The controlled mechanical manipulation of individual molecules with a scanning probe microscope (SPM) allows the fabrication of single-molecule devices [1,2] and metastable supramolecular assemblies [3]. Machine learning can reduce the tedious work of finding a successful manipulation protocol for a certain manipulation tasks. We use reinforcement learning (RL) to automatically solve the prototypical task of removing a single PTCDA (perylene-tetracarboxylic dianhydride) molecule from a hydrogen-bonded assembly [3]. Since our RL application is not fully in-silico but receives its feedback from an actual experiment, it needs a training efficiency on par with a human to be useful. We achieve this by teaching the machine some “intuition” about the Cartesian space in which the manipulation takes place by, e.g., spawning a series of weak learners along all tried trajectories and training on unseen state-action pairs. Our method could be a blueprint for solving various manipulation tasks posed in a Cartesian space.
C. Wagner et al. Phys. Rev. Lett. 115, 026101 (2015)
T. Esat et al. Nature 558, 573 (2018)
M. F. B. Green et al. Beilstein J. Nanotechnol. 5, 1926 (2014)