Deep-learning applied to coherent combining of fiber lasers: toward an experimental demonstration in a non-controlled environment

Coherent beam combining (CBC) by active phase control has strong potential to go beyond current fiber laser limitations for power scaling by coherently adding the power of multiple emitters. However current methods are limited by the electronics’ bandwidth. In simulated environments, deep-learning techniques have proven their potential to improve the bandwidth of phase control by compensating the phase errors in one step while maintaining excellent CBC efficiency. To work toward an experimental application, we present in this paper a new tool to optimize the imaging camera position. We also present another new tool to design and test neural networks for CBC by comparing training duration and size of the training-data space on CBC performance. This tool can also be used to predict CBC speed and efficiency in an active phase loop allowing to tailor the neural network to fit CBC specifications. Finally, we observe a mismatch between real experimental interference pattern images and simulations. This mismatch might be due to defects such as tip-tilt or other beam aberrations. Since then, we have significantly improved the similarity between experiment and simulation and managed to perform a successful experimental test of AI-enhanced CBC.