Applying artificial neural networks to coherent control experiments: A theoretical proof of concept

Abstract

We propose a method of experimental coherent control that exploits partial and/or prior knowledge of a molecular system to efficiently arrive at a solution by using an artificial neural network (ANN) to generate a control field in consecutive temporal steps based on dynamic experimental feedback. Using a one-dimensional double-well potential model corresponding to the torsional motion of 3,5-difluoro-${3}^{\ensuremath{'}},{5}^{\ensuremath{'}}$-dibromobiphenyl (${\mathrm{F}}_{2}{\mathrm{H}}_{3}{\mathrm{C}}_{6}\ensuremath{-}{\mathrm{C}}_{6}{\mathrm{H}}_{3}{\mathrm{Br}}_{2}$) to outline and verify our approach, we theoretically demonstrate that an optimized ANN can achieve robust quantum control of nuclear wave-packet transfer between wells despite the addition of random perturbations to the simulated molecular potential energy and polarizability surfaces. We suggest that under certain conditions this will also allow the ANN to achieve the stated control objective in an experimental situation. We show that the number of measurements our method requires to generate an optimized field is equal to the dimensionality of the optimization problem, which is significantly less than a naive closed-loop approach would generally need to achieve the same results.