Analog logic gate of structured light based on nonlinear optical neural networks

Optical Neural Networks combine the typical architecture of a neural network, with the advantages of using light to store the information and perform the operations. They have multiple applications like image recognition or optical computing. Nonlinearity can be added to these systems by having layers of materials with a nonlinear response in their interaction with light. This project aims to use a nonlinear optical neural network as basis for an all optical decoder, showing the potential of these systems.

Moreover, this work focuses on designing and experimental implementing the physical neural network. It starts with a study on the Kerr nonlinearity of selected materials in order to choose the best. Such study required performing z-scans and I-scans on the samples. Through these measures, the best sample to use in the experiment was found to be a gold bearing glass.

Then, the focus was to improve a machine learning model to act as a digital twin to the real system. This model served to test system architectures and will be responsible for training the parameters later fed into the physical neural network. Following this, a real experimental setup was designed, assembled and aligned. The latter presented several challenges that had to be overcome. As a first approach, a linear model was chosen for testing. The model was trained in silico and then tested in the real setup, and such results are presented. The work here presented provides a ground base for the experimental implementation of an all optical decoder.