Estimation of electromagnetic model of light in different media using artificial neural network

In this thesis study, the design of a power splitter, a type of optical waveguide, was implemented using a generative adversarial network structure based on deep learning methods within artificial intelligence instead of numeric methods. The designed power splitter has a design area of 2.5 x 2.5 μm² and is features high resolution (200 x 200 pixels). The design of this high-resolution power splitter structure was accomplished for the first time using a deep learning method.The study focuses on chip-level waveguides produced using deep learning methods, highlighting their efficiency, high resolution, and asymmetric power distribution. The optical structure, comprised of one input and two outputs, exhibits the characteristics feature of distributing electromagnetic power asymmetrically. The data sets provided to the deep learning structure, specifically the DCGAN (Deep Convolutional Generative Adversarial Network), were curated by selecting high-efficiency optical waveguides from those initially randomly generated in the design area. The design area of the optical waveguide, featuring a binary pixel distribution of silicon(1) and air(0), plays a crucial role in achieving the desired outcomes. The results indicate that the power splitter structure, generated by artificial intelligence, efficiently transmits electromagnetic wave energy with low loss and reflection rates. The efficiency ratios of the produced power splitter structures generally fall within the range of 85% to 95%, with the highest efficiency for power splitter having 20 x 20 pixel design area reaching 97.9%, for power splitter having 200 x 200 pixel design area reaching an impressive 95.9%.