Sprecher
Beschreibung
Laser beam alignment is a non-trivial and time-consuming problem native to a multitude of present-day experiments. We introduce a reinforcement learning-based laser beam alignment system that learns to align a Mach-Zehnder interferometer and an off-axis parabolic mirror with live optimization correcting for beam drift or externally introduced mirror misalignment. The algorithm manages to find a technique for recovering a beam lost from its field of vision. This technique allows for the use of open-loop motors as the agent is agnostic of the orientation of the component it controls.
We will explore different algorithms such as DQNs, Actor-Critic, SARSA and show their respective advantages and introduce a new way of simulating beam propagation to train focal spot optimisation.