Deep learning (DL) has been applied to phase control in coherent beam combining (CBC) recently. However, existing DL-based approaches for filled-aperture CBC essentially convert the phase-locking path into tiled-aperture schemes. Consequently, common-path phase locking in DL-based filled-aperture CBC remains unrealized. Common-path refers to a phase-locking scheme in which the phase information is extracted from the combined beam after the same combining system. In this paper, a common-path phase-locking method is proposed. By exploiting the intrinsic nonuniformity, each laser source is effectively labeled, enabling a mapping between the combined speckle and the multi-source phase. A neural network is employed to reconstruct the phase. Simulations with 25-channel CBC demonstrate a phase-locking accuracy of up to λ/39. Notably, it remains effective under dynamic phase disturbances. Our work presents a common-path phase-locking approach based on a neural network for filled-aperture CBC, which can offer a new solution for the field.
