An efficient neural network for predicting supercontinuum dynamics in photonic crystal fiber

This study investigates the generation of supercontinuum (SC) in photonic crystal fibers (PCFs) using chalcogenide materials. For this purpose, generation and translation-based convolutional neural network (CNN) models, Gen-Net and Trans-Net, are developed to predict the spatial and temporal evolution of SC generation, based on a dataset of multiple materials with varying optical properties. Instead of relying on spectral power data across different wavelengths, spatial and temporal evolution plots are treated as images, which reduces the dataset size required for simulations. This approach allowed for efficient handling of large datasets, making it a scalable model. The dataset includes four chalcogenide materials: Ge_11.5As₂₄Se_64.5, As₂S₅, As₂Se₃, and As₂S₃, and the model learns the relationship between pulse evolution and material optical properties. The unified dataset comprises supercontinuum data for the four materials, demonstrating the model's ability to produce accurate pulse propagation responses under identical input conditions. The two models function as a two-stage platform for modeling pulse propagation. The models' performance was evaluated using Average Pixel Deviation (APD), yielding values close to 0.03, indicating high precision in capturing nonlinear dynamics. Furthermore, the models achieved computational speeds nearly 330 times faster than numerical methods, marking a significant advance in modeling nonlinear optics and offering valuable insights for future research in photonic device applications.