Study of Linear and Non Linear Properties of Photonic Structures via Numerical and Deep Learning Techniques

Abstract

This study focuses on advancing the design, characterization, and optimization of photonic structures by integrating elementary and advanced deep learning tech niques. The primary aim is to develop predictive models that enhance both the accuracy and efficiency of photonic structure simulations, enabling faster design it erations and real-time optimization.This is aimed to substitute the conventionally used simulation softawares for similar purpose.The first part involves utilizing Ar tificial Neural Networks (ANNs) to predict key linear optical parameters of PCFs, including effective refractive index, dispersion, mode purity, and effective area. The ANN model achieved a high prediction accuracy, evidenced by a Mean Squared Error (MSE) of 0.0218, with a training time of approximately 31 seconds. This rep resents a substantial reduction in computational time compared to traditional nu merical methods while maintaining precise predictions, demonstrating the model’s practical applicability.The second objective focuses on modeling the nonlinear dy namics of supercontinuum generation in optical fibers, with particular attention to chalcogenide planar waveguides. An ANN-based deep learning model was em ployed to accurately capture the complex pulse propagation dynamics, achieving an MSE of 0.0032 with a computational training time of 472 seconds. The model’s abil ity to predict spectral evolution across various pump powers accurately provides a significant potential for real-time optimization of experimental setups.The third objective introduces a novel approach using Closed-Form Continuous-Time Neural Networks (CfC) for predicting supercontinuum spectra in waveguides with diverse material compositions. The CfC model outperformed conventional methods such as Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) net works, achieving high accuracy with MSE values as low as 7.4199 × 10−4 for a three material dataset and 6.9408 × 10−4 for a four-material dataset. The model demon strated low computation time, confirming its robustness, scalability, and adaptabil ity across different materials.Overall, this research demonstrates that integrating machine learning techniques with photonics significantly enhances the accuracy of predictive models while reducing computational time. These advancements have practical implications for the design, optimization, and real-time control of photonic devices, paving the way for future innovations in the field.