Deep neural network for microstructured polymer fiber modeling

Abstract

Abstract Microstructured polymer fibers have been widely studied for terahertz waveguides, sensing, environmental monitoring, and medicine. Time-consuming methods, including the finite element method and finite-difference time-domain, have been used to design and simulate microstructure polymer fibers. Designing optimal artificial neural networks (ANN) with various hyperparameters for each optical property considerably improve the prediction accuracy. We compared and explained the differences between predicting one quantity at one time and predicting multiple quantities at the same time. We utilized the successfully trained ANN to accurately evaluate the optical properties of the unknown geometry, including the effective refractive index ( n eff ), and effective mode area ( A eff ). In addition, we took logarithmic transformation to form a more uniform distribution, which allowed us to successfully predict confinement loss (α c). We also demonstrated that ANN could predict the core mode properties of unknown geometric parameters in a considerably short time compared to the finite element method simulation. Moreover, ANN with ultrafast calculation speed and high prediction accuracy accelerate the optical function device design process.