The optical neuron model presented in this master’s thesis is based on third-order nonlinear optical phenomena (the Kerr effect and third-harmonic generation). In the introduction, I outline the theoretical framework that connects nonlinear optics with the fundamental building blocks of artificial neural networks. The main part of the work focuses on the numerical implementation of one-dimensional FDTD simulations used to study the propagation of electromagnetic waves through a Kerr medium and the generation of the 3$\omega_0$ frequency component. From the dependence of the third-harmonic amplitude on the input amplitude, I derived a photonic activation function; the logistic function showed the best fit and was therefore used in further simulations. I then constructed a photonic neural network that incorporates key physical constraints (non-negative weights and activation saturation) and compared its performance on the MNIST dataset with reference electronic networks. The results show that photonic constraints combined with the extended activation enable high accuracy (approximately 95–97 %), while electronic networks achieve slightly higher accuracy (above 97–98 %). This confirms that photonic neural networks are a promising candidate for replacing electronic neural networks. Proposed future directions include extending the simulation of light propagation to tree dimensions, exploring different materials, and conducting experimental validation.
|
