Generative Adversarial Neural Networks Model of Photonic Crystal Fiber Based Surface Plasmon Resonance Sensor

Photonic crystal fibers (PCF) for specific applications are designed and optimized by both industry experts and researchers. However, the potential number of design combinations and flexible propagation features are offering a wide range of application areas, resulting in an exponential increase in the search space. This issue combined with the speed of the commonly used full vectorial finite element method (FV-FEM) simulators causes the task to take a significant amount of time. As stated in the previous works, artificial neural networks (ANN) can be employed to predict the result of numerical simulations much faster. However, there are two issues with the methods proposed previously: the required number of samples for training and the generality of these methods. In this article, we propose the use of generative adversarial networks (GAN) to augment the real dataset to train an ANN model. The experimental analysis suggested that the proposed combination not only accurately predicts the confinement loss even with limited amounts of data but also GAN can be used to improve existing methods in the literature. Moreover, the proposed system can also predict the confinement loss over a range of analytes and wavelengths in a completely new geometric configuration and generic enough not to require any additional tuning when used in a new dataset. This property is demonstrated by experimenting on an existing PCF dataset, which the proposed system surpassed the accuracy of the method proposed for it.