Aprendizado de máquina aplicada em modelagem da eficiência de acoplamento entre guias dielétricos.

Abstract

This dissertation aims to use machine learning for the formulation and design of mimetic models of TAPER-type photonic devices intended for coupling waveguides with different geometric structures. The objective is to evaluate the coupling efficiency as a function of specific variations in geometric characteristics in the C-Band. In addition to the theoretical foundation, it was necessary to prepare a database to train the learning networks. This database consists of numerical solutions obtained through a finite element-based numerical method, as well as previously published information. These data were consolidated into a comprehensive set whose attributes correspond to variations in the dimensions of the taper segments (the length denoted as “a”), and the output is the coupling efficiency (represented by “η”). A neural network architecture was then developed with the input parameters: length of each of the 15 taper segments (a), wavelength (λ), refractive indices of the core (n1) and substrate (n2), and as the output parameter: the ratio between the input power (Pin) and the output power (Pout) given by the coupling efficiency (η). For this architecture, variations in training algorithms and activation functions were explored. These variations were used to evaluate the performance of the proposed models, considering criteria such as accuracy, precision, simplicity, and the computational costs involved. As a result, the developed architectures demonstrated performances better than the values defined by the stopping criteria, with a mean squared error less than 10-7 and a regression rate or determination coefficient R2 of 100% in more than 92% of the total of 81 models evaluated with reduced use of computational resources. This study aims to contribute to the improvement of the understanding and design of photonic devices through the synergistic application of machine learning and traditional techniques.