Optical properties prediction of negative dispersion-compensating photonic crystal fiber using machine learning

Abstract

Dispersion in optical fiber communication causes signal degradation as light pulses spread out over distance. Dispersion compensation techniques, such as dispersion- compensating fibers, fiber Bragg gratings, and chirped fiber gratings, counteract this effect. Photonic Crystal Fibers (PCFs) offer a promising solution for dispersion com- pensation by counterbalancing the positive dispersion of standard fibers. PCFs have emerged to be a dynamic and adaptable solution for various applications such as en- compassing spectroscopy, bio-sensing, metrology, and long-haul optical communica- tion systems, owing to their remarkable dispersion-compensating characteristics. In this study, a novel and highly efficient Negative Dispersion-Compensating Photonic Crystal Fiber (NDC-PCF) is designed, and machine learning approaches, specifically Random Forest and Artificial Neural Network (ANN) models, are proposed for predicting output properties, including effective refractive index, dispersion, confinement loss, effective area, and V-parameter. The machine learning models are trained and tested on input parameters within a wave- length range of 1.18–1.75 µm, pitch from 0.75–0.9 µm, core diameter, and air holes in the cladding region. Settling on a pitch value of 0.8415 µm achieving of a minimal dispersion of -1582.21 ps/(nm-km), critical for efficient optical fiber communication. Compared to conventional numerical simulations such as COMSOL Multiphysics and ANN, the proposed Random Forest model demonstrates significantly reduced computa- tional resources and time requirements, with training in milliseconds and testing in less than one millisecond. The model achieves an impressive average accuracy of around 98%. On the other hand, the best ANN model is obtained by fine-tuning hyperparam- eters, such as the count of hidden layers, nodes, and training iterations, and attains an average accuracy of approximately 99.99% for all parameters, with training taking several seconds and prediction in milliseconds. With its negative characteristics, the proposed NDC-PCF exhibits great potential for real-world applications in high-speed optical communication networks. The utilisation of machine learning approaches offers an efficient alternative to conventional numerical simulations for predicting the optical properties of NDC-PCFs, enabling faster design optimisation and cost reductions.