Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 42: Focus Session: Innovation in Machine learning PRocEsses for Surface Science (IMPRESS)

O 42.2: Hauptvortrag

Dienstag, 17. März 2020, 11:00–11:30, TRE Phy

Machine learning for molecular nanorobotics — •Christian Wagner — Peter Grünberg Institut (PGI-3), Forschungszentrum Jülich, Germany — JARA Fundamentals of Future Information Technology, Jülich, Germany

The ability to handle single molecules as effectively as macroscopic building-blocks would enable the construction of complex supramolecular structures that are not accessible by self-assembly. A central difficulty on the way towards this technology is the uncontrolled variability and poor observability of atomic-scale conformations, especially during the manipulation process. We present a generic strategy to overcome both obstacles, and demonstrate autonomous nanorobotics with single molecules using reinforcement learning (RL). Quite generally, RL is able to learn strategies even in the face of large uncertainty and with sparse feedback. Indeed, RL based prediction models recently exceeded human performance in several games. However, to be useful for nanorobotics, standard RL algorithms must be adapted to also cope with the limited training opportunities that are available there. We demonstrate our correspondingly enhanced RL approach by applying it to an exemplary task of subtractive manufacturing with a scanning probe microscope (SPM). Complementary to that we outline how machine learning and control theory methods in combination with molecular simulations can be utilized to recover atomic-scale conformations from the sparse experimental SPM data available during manipulation.