Machine-learning autonomous molecular manipulations

Bernhard Ramsauer1, Grant J Simpson2, Leonhard Grill2, Oliver T Hofmann1

1Institute of Solid State Physics, TU Graz, Austria
2Institute of Chemistry, Uni Graz, Austria

Finding the optimal parameters to manipulate individual molecules with STM is challenging and time consuming, even for human experts. This is particularly the case for larger molecules, where there molecule cannot be moved with arbitrary precision and the final location after a manipulation attempt is given by a probability distribution. Here, we present a combination of a reinforcement-leaning based exploration strategy with Bayesian statistics to find efficient protocols that allow bring molecules to arbitrary positions on the surface and to assemble them into functional nanostructures. Exemplarily, we build a “tic-tac-toe” board and a quantum corral. In both cases, the automated assembly achieves the assembly much faster than a human operator. Still, the remaining time bottleneck is the necessity to (re-)locate the molecule after each manipulation. We discuss options to overcome this via automated image recognition, neural networks to analyze the tunneling current during manipulation, and other search strategies.