Reinforcement learning for tiled aperture beam combining (Conference Presentation)

We developed a simulation of a beam tiled aperture combining system with n emitters. Using the Fourier propagation method, we calculate the far-field pattern depending on (n-1) relative phase parameters. We then simulate the phase noise as a time series using Gaussian random walk phase noise, with the possibility to actuate the phase parameters. For the stabilization scheme, we exclusively use a far filed intensity pattern. This image we feed through a simple two-layer neural network trained by a deep deterministic gradient reinforcement learning algorithm. This allows for a large amount of flexibility in terms of optimization metrics, which easily allows for beam combination in the far filed and beam shaping. The control policy is automatically derived from reward functions which can be phase patterns, target beam shapes or maximum power through a pinhole without the need for further optical analysis. After training only the image is used for stabilization.