Confinement loss prediction in diverse anti-resonant fibers through neural networks.

In this work, genetic algorithm (GA) is employed to optimize convolutional neural networks (CNNs) for predicting the confinement loss (CL) in anti-resonant fibers (ARFs), achieving a prediction accuracy of CL magnitude reached 90.6%, which, to the best of our knowledge, represents the highest accuracy to date and marks the first instance of using a single model to predict CL across diverse ARF structures. Different from the previous definition of ARF structures with parameter groups, we use anchor points to describe these structures, thus eliminating the differences in expression among them. This improvement allows the model to gain insight into the specific structural characteristics, thereby enhancing its generalization capabilities. Furthermore, we demonstrate a particle swarm optimization algorithm (PSO), driven by our model, for the design of ARFs, validating the model's robust predictive accuracy and versatility. Compared with the calculation of CL by finite element method (FEM), this model significantly reduces the cost time, and provides a speed-up method in fiber design driven by numerical calculation.