Reverse engineering and analysis of microstructure polymer fiber via artificial neural networks: simplifying the design approach

Abstract

Abstract Microstructure polymer fibers have been extensively researched for their applications in various fields. The design and simulation of these fibers have utilized time-consuming techniques like the finite-difference time-domain and finite element method to facilitate the process. This study proposes an optimal artificial neural network (ANN) for predicting the structural design required to achieve desired optical properties. The ANN model takes various optical properties, including confinement loss, effective index, effective mode area, and wavelengths, as inputs to predict fiber design parameters such as diameter and pitch. To address the challenge of skewed distributions, a data set with a Gaussian-like distribution for confinement loss was generated using a logarithmic transformation method, enabling effective training of machine learning models. Furthermore, the ANN model demonstrates its capability to rapidly predict unknown geometric parameters using only the core mode properties of a polymer fiber, achieving results in a significantly shorter time (3 ms) compared to the trial-and-error approach of finite element method simulation (15 s). The reverse engineering model achieves a mean square error of 3.4877 × 10−06 with five hidden layers. The ANN model not only offers ultrafast calculation speed but also delivers high prediction accuracy, thereby accelerating the design process of optical devices. The differentiation among the prediction result, target, and calculation result provides compelling evidence that the proposed approach is an effective methodology for designing microstructure polymer fibers.