GS-RNN: a phase-only hologram generation model based on the fusion of the Gerchberg-Saxton algorithm and a neural network

In this paper, we propose a computer-generated hologram (CGH) iterative generation model based with a recurrent neural network architecture, which combines an improved residual complementary neural network with the traditional Gerchberg-Saxton (GS) algorithm. This model is referred to as GS-RNN and is designed to generate high-quality phase-only holograms (POH) with a reduced number of iterations. By incorporating trainable parameters from the neural network into the GS algorithm, GS-RNN enhances the model's performance through learning from sample data. After learning from the samples, the iterative computation efficiency of GS-RNN achieved three times that of the traditional GS algorithm. Even when the physical parameters for generating POH are changed, GS-RNN can maintain its outstanding performance without the need for retraining. Furthermore, the optical experimental results are consistent with the simulation results, further validating the effectiveness and success of the proposed GS-RNN model.