Design of binary blazed grating couplers based on cascaded residual neural networks

Focusing on improving the inverse design method of binary blazed grating couplers to achieve high design efficiency and low error, this work investigates three approaches: a cascaded neural network, an improved cascaded residual network, and particle swarm optimization (PSO). Firstly, a comprehensive training dataset was obtained through electromagnetic simulations, and model hyperparameters were determined. The ordinary cascaded neural network required a training time of 33,944 s, achieving a relative error of 0.12. To enhance both design efficiency and accuracy, an improved cascaded residual neural network model was developed. By introducing residual connections, it effectively mitigated issues of gradient vanishing and gradient explosion. With this improvement, the training time was significantly reduced to 3737 s, and the relative error was lowered to 0.08. Additionally, PSO was applied to optimize the grating coupler. Using a population size of 50 and performing 100 iterations, the optimization process required approximately 45,000 s, a maximum coupling efficiency of 47.37%. A comprehensive comparison of these methods demonstrates that the improved cascaded residual network exhibits significant advantages in both training time and relative error. This highlights its great potential for significantly improving the inverse design efficiency and accuracy of binary blazed grating couplers.