Prediction of fundamental modes in ridge waveguides with convolutional neural networks

Control and manipulation of light are crucial topics in optics and photonics. Ridge waveguides are a type of rectangular waveguide designed to provide effective control over light. These customized designs stand out due to their high mode confinement, efficient optical communication with low losses, and flexibility for different applications. The mode profile supported by the waveguide, which is determined by the ridge width and height, is generally calculated using numerical methods such as Finite Difference Eigenmode (FDE) and Eigenmode Expansion (EME). However, as the structures become more complex, the computational costs associated with these methods increase significantly. A ridge waveguide using a 0.5 mu m wide and 0.18 mu m high silicon core on a silicon dioxide substrate was designed in this study. The fundamental mode with transverse electric polarization was analyzed using the FDE method. The fundamental mode profiles derived from varying ridge heights, widths, and light wavelengths were organized to provide the dataset for the training of the transposed convolutional neural network (CNN) designed for mode estimation. The efficiency of this transposed-based CNN model in mode profile estimation was assessed by the computation of several learning performance measures, including MAE, MSE, and RMSE. The results of this study demonstrate the capability of the created deep learning model to serve as an alternative to conventional approaches in computational electromagnetic applications.