Algoritmo baseado em machine learning para projeto de dispositivos de interferência multimodo

Abstract

The improvement of Photonic Integrated Circuits to make them more flexible, reconfigurable and compact has been the norm for telecommunications systems. Against this backdrop, attention has turned to Multimode Interference devices due to the advantages presented by their properties and functionalities. This dissertation project proposes using Machine Learning techniques for the inverse design of Multimode Interference devices, as a power divider in the wavelength range of 1.25 ~ 1.7µm integrating the O, E, S, C, L and U bands. To do this, in addition to carrying out a literature survey, it was necessary to simulate Multimode interference devices in specific software and develop an algorithm using Artificial Neural Networks. The neural network architecture was configured with the following input parameters: wavelength (λ), core refractive index (n1), substrate refractive index (n2), device width (WMMI) and transmission efficiency (%) and as output parameters it targeted the coordinates of the x-axis1, x-axis2 referring to the position of the output ports (1 × M) of the device and the coordinates of the y-axis referring to the cut-off length (LMMI) of the arrangement for the highest coupled power of the output ports. As a result, the Artificial Neural Network developed presented a cross-validation Mean Square Error equivalent to 6.39410 × 10-5, a linear regression of 0.99997 and a computational processing time of 7.46 seconds, capable of providing data for the design of compact Multimodal Interference devices with dimensions from 2.00 × 5.32µm and losses ranging from 0.32 to 0.47dB for the most efficient device.