Coherent beam combination via deep learning using differential mapping

Coherent beam combination (CBC) is a route to breaking through the power limit of single fiber laser amplifiers, furthermore, tiled phased arrays enable rapid, flexible beam shaping for optimizing laser-based processes. In this work, we present PCANN, a principal component analysis (PCA) and neural network (NN) algorithm to simultaneously phase lock and beam shape CBC arrays, with no knowledge of phase, which we experimentally demonstrate with a three element tiled array from ytterbium-doped fiber amplifiers. The PCANN uses a differential mapping scheme through applying arbitrarily scaled phase controller inputs and recording pre- and post-step beam intensity images, allowing labeled training pairs to be generated without requiring direct phase measurement. The PCA reduces 64×64 pixel images to 13 features, enabling a lightweight mixture density NN to provide phase-actuator voltage corrections at 1.1 kHz with 1.4 ms latency. It achieved λ /150 (41 mrad) RMS residual phase error, single-step beam shaping in 1 ms, and robustness over days. Scalability to higher channel counts is shown with a seven-channel fiber CBC experiment, and simulations up to 61 fibers indicate linear scalability in training data requirements.