Redefinable neural network for structured light array

Abstract

Neural networks have provided faster and more straightforward solutions for laser modulation. However, their effectiveness when facing diverse structured lights and various output resolutions remains vulnerable because of the specialized end-to-end training and static model. Here, we propose a redefinable neural network (RediNet), realizing customized modulation on diverse structured light arrays through a single general approach. The network input format features a redefinable dimension designation, which ensures RediNet wide applicability and removes the burden of processing pixel-wise light distributions. The prowess of originally generating arbitrary-resolution holograms with a fixed network is first demonstrated. The versatility is showcased in the generation of 2D/3D foci arrays, Bessel and Airy beam arrays, (perfect) vortex beam arrays, and even snowflake-intensity arrays with arbitrarily built phase functions. A standout application is producing multichannel compound vortex beams, where RediNet empowers a spatial light modulator (SLM) to offer comprehensive multiplexing functionalities for free-space optical communication. Moreover, RediNet has the hitherto highest efficiency, only consuming 12 ms (faster than the mainstream SLM framerate of 60 Hz) for a 1000 2 -resolution holograph, which is critical in real-time required scenarios. Considering the fine resolution, high speed, and unprecedented universality, RediNet can serve extensive applications, such as next-generation optical communication, parallel laser direct writing, and optical traps.