Prediction of the optical properties in photonic crystal fiber using support vector machine based on radial basis functions

Abstract

Compared with traditional optical fiber, photonic crystal fiber (PCF) has many novel optical properties owing to its diversity in cladding distribution. But, it is a problem to measure the optical properties of PCF under variable structural parameters. Artificial neural networks (ANNs) have been used to predict the optical properties of PCF, but ANNs have the multiple local minima problems, whereas support vector machines’ solutions are globally unique and will not fall into local minimum values. In this paper, support vector machines (SVMs) based on radial basis functions were used to predict the effective refractive index (neff), chromatic dispersion (D), and confinement loss (αc) of PCF. Well-trained SVMs can accurately and quickly predict the optical properties of any geometric parameters in the studied parameter space. A data set similar to Gaussian distribution was formed by two logarithmic transformation methods to avoid the problem that machine learning models cannot be well trained on extremely skewed distribution. Compared to ANNs, SVMs are more accurate and show stable prediction results.