Nian Wu1, Markus Aapro1, Alexander Ilin2, Robert Drost1, Joakim Jestilä1, Nan Cao1, Roman Fasel3, 4, Peter Lijeroth1 and Adam S. Foster1,6
1Applied Physics, Aalto University, Espoo, Finland
2Computer Science, Aalto University, Espoo, Finland
3Nanotech@surfaces Laboratory, Empa-Swiss Federal Laboratories for Materials Science and Technology, 8600 Dübendorf, Switzerland
4Department of Chemistry, Biochemistry and Pharmaceutical Sciences, University of Bern, 3012 Bern, Switzerland
5Departamento de Química Orgánica, Universidad Autónoma de Madrid, Campus de Cantoblanco, 28049 Ma-drid, Spain
6WPI Nano Life Science Institute, Kanazawa University, Kanazawa, Japan
Several breakthrough studies have harnessed scanning probe microscopy (SPM) manipulations to control chemical reactions in on-surface molecular synthesis. In general, for scanning tunnelling microscope (STM) manipulations, they are predominantly controlled via parameters of the tip position, pulse voltages and tunneling conductance. However, the selection of proper parameters requires extensive domain knowledge, which is time consuming and not necessarily transferable to new systems. Recent research has allowed the automation of a wide range of challenges in SPM, including image quality assessment, lateral and vertical manipulation. However, the automation for breaking or forming covalent bonds, which is an indispensable step during chemical synthesis is, as yet, unexplored. To address this problem, we developed our deep reinforcement learning (DRL) approach to automate bromine removal from 5,15-bis(4-bromo-2,6-methyl-phenyl)porphyrin (Br2Me4DPP) through learning manipulation parameters in STM. Meanwhile, DFT calculations were implemented to explore the reaction mechanism in combination with STM results and DRL results.