Efficient phase identification in coherent beam combination using interpretable deep learning

Coherent beam combination (CBC) offers significant power scaling beyond the capabilities of individual fiber lasers, but its effectiveness is heavily reliant on precise phase stabilization. Recent advancements in deep learning have shown potential for phase retrieval from interference intensity patterns in a single step. However, the interpretability of deep learning models and the optimal positioning of the imaging system remain unresolved challenges. In this study, we employ a spatial light modulator to emulate a CBC system and systematically investigate the phase prediction accuracy at various axial positions of the imaging system. We demonstrate that phase-retrieval efficiency can be substantially enhanced by identifying regions of the interference pattern with higher phase sensitivity. This approach enables a significant reduction in input size, allowing for the use of a lightweight fully connected neural network to achieve a phase prediction error of approximately λ/60, with a pure fully connected neural network inference rate of approximately 35 kHz for a 7-beamlet hexagonal close-packed array.