Neural architecture search generated phase retrieval network and its application in laser wireless power transmission

Laser wireless power transmission (LWPT) faces critical challenges from atmospheric turbulence, and deep learning-based adaptive optics (AO) systems are a promising solution. However, the architectures of existing phase retrieval neural networks are all empirically designed, which limits the performance of AO systems in complex and variable LWPT links. In this paper, a neural architecture search-generated phase retrieval network (NAS-PRNet) is proposed, which jointly optimizes both the weight and architecture parameters, thereby improving the wavefront compensation accuracy and environmental adaptability of the AO system. Simulations and LWPT experiments based on a 40-element AO system are conducted to validate the proposed method. Simulation results show that the performance of the NAS-PRNet with a fully automatically searched architecture can approach or even surpass that of the empirically designed network, achieving an 8.43% reduction in root mean square error (RMSE) of Zernike coefficient prediction. Experimental results demonstrate that under closed-loop control, NAS-PRNet increases the average output power of the photovoltaic array (PVA) by 105.07% and reduces the power fluctuation by 94.06%, significantly improving both transmission efficiency and stability. This work provides an effective, automated solution for the mitigation of turbulence effects in complex LWPT environments.