Machine learning methods for estimating bent photonic crystal fiber based SPR sensor properties

In the recent years, the use of machine learning approaches in optical devices and fibers is increasing. However, most methods concentrate on the use of Artificial Neural Network (ANN) methods due to the ability of automatically fitting to the problem. In this work, a classical non-linear regression method, namely k-Nearest Neighbor Regression (KNNR) is proposed for determining the loss characteristics of a photonic crystal fiber (PCF) based surface plasmon resonance (SPR) sensor in the presence of a bend in either x or y direction. Although KNNR is a simple method, it is very well known that in certain systems it can out-perform ANN. It is believed that PCF based structures can be a good candidate for this comparison. In order to judge the performance of different regression techniques, we have built a database that contains 1180 samples. The dataset contains PCF structure data for non-bent(straight fiber), bent in x and y-directions. Experiments show that KNNR outperforms both ANN and Linear Least Square Regression methods even when a feature space expansion method is employed. In addition, KNNR does not require any lengthy training process, allowing it to be used instantly once the training data is available. This can be exploited to complement existing simulation techniques.