Deep Learning-Optimized Design of Asymmetric Optical Directional Couplers for High-Efficiency Coupling

Abstract

This study investigates the application of a deep neural network to optimize the design of asymmetric optical directional couplers, addressing the limitations of traditional iterative methods, which are often resource-intensive and time-consuming. A feed-forward neural network was developed to predict the coupling length of a coupler based on its structural parameters. A large dataset of 30,000 samples was used to train, validate, and test the model. The network achieved a minimal mean square error (MSE) of 1.07×10−6, demonstrating high precision. The trained model exhibited a remarkable prediction accuracy of 99.88% for the coupling length. To validate these results, finite-difference time-domain (FDTD) simulations were conducted using the network's predictions. The simulations confirmed that injecting a Transverse Electric fundamental mode (TE0) into the access waveguide successfully excited a double mode (TE1) in the bus waveguide, achieving a high coupling efficiency of 99.5%. This research demonstrates that a deep learning approach can provide a rapid and accurate method for designing efficient mode-conversion-based optical asymmetric directional couplers.