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

Area of Science:

• Nonlinear Optics
• Materials Science
• Computational Physics

Background:

• Supercontinuum (SC) generation in photonic crystal fibers (PCFs) is crucial for various photonic applications.
• Modeling SC generation traditionally involves computationally intensive numerical simulations.
• Chalcogenide materials offer unique optical properties for SC generation.

Purpose of the Study:

• To develop accurate and efficient models for predicting SC generation in PCFs.
• To investigate the use of convolutional neural network (CNN) models for simulating pulse propagation.
• To explore the application of these models across different chalcogenide materials.

Main Methods:

• Development of two CNN models, Gen-Net and Trans-Net, for predicting spatial and temporal SC evolution.
• Treatment of SC evolution plots as images to reduce dataset size and enable scalable modeling.
• Training the models on a unified dataset of four chalcogenide materials (Ge_11.5As₂₄Se_64.5, As₂S₅, As₂Se₃, As₂S₃).

Main Results:

• The CNN models accurately predict pulse propagation responses, achieving an Average Pixel Deviation (APD) close to 0.03.
• The models demonstrated scalability and efficient handling of large datasets.
• Computational speeds were approximately 330 times faster than conventional numerical methods.

Conclusions:

• Gen-Net and Trans-Net provide a highly efficient and accurate two-stage platform for modeling nonlinear pulse propagation.
• This AI-driven approach significantly advances the simulation of nonlinear optics.
• The findings offer valuable insights for the design and application of photonic devices.