Learning and control for molecular nanofabrication
To advance the field of SPM-based fabrication of molecular nanostructures, we introduce machine learning and control theory to the field. For example, we pioneered autonomous nanofabrication using a reinforcement learning agent and combined machine learning and control theory both to reveal molecular conformations during manipulation and to speed up scanning quantum dot microscopy.
Our goal is to understand and utilise molecular manipulation [1]. The complexity of manipulation experiments and the corresponding high computational cost for realistic simulations calls for a data-driven approach. We follow a multidimensional strategy. On the one hand, we focus on molecular fabrication by a self-learning agent, thereby turning the SPM into an autonomous robot capable of achieving certain manipulation tasks [2]. The agent is designed to cope with the sparse information available from the SPM and with the intrinsic variability at the nanoscale. Importantly, the agent does not need to know the actual molecular conformation.
On the other hand, we utilise machine learning and concepts of control theory, such as particle filters, to “observe” molecular conformations in the experiment. To do so, machine learning is, for example, used to speed up molecular simulations [3]. When complemented with virtual reality [4], these concepts will ultimately allow us to visualise changes in the molecular conformation in real time during a manipulation experiment.
