From Optimization to Computation : Metaheuristic Control of Optical Fiber Systems and Optical Computing

Abstract

The propagation of optical pulses in fibers exhibits complex nonlinear dynamics, especially in the femtosecond regime, where spectral features are highly sensitive to input parameters. With growing demand for specialized light sources, controlling these features is crucial. However, the interaction between pulse input parameters and resulting features after its nonlinear propagation is non-trivial. Machine learning has emerged as a powerful tool for creating new functionality and improving the performance of many complex systems, without needing precise models. ln this thesis, we combine machine learning algorithms with nonlinear propagation of short pulses in two ways: (i) we leverage advanced numerical algorithms for systematic control and optimization of nonlinear pulse propagation. (ii) We explore the use of optical fibers as platforms for computing.

We implement a rapid and flexible genetic algorithm on various fiber systems, demonstrating its capability for self-alignment and drive-resilience. First, we control fiber instabilities from a noise-like pulse cavity, reaching a stable state in just a few minutes. Second, we apply the same principle to the coupling of orbital angular momentum modes to a fiber, stabilizing the system for over a day. Third, we leverage several metaheuristic algorithms for the spectral shaping of supercontinuum gener­ation, controlling up to four wavelengths at the same time. We use simulations to guide the implementation of the experiments. We then compare their performances in order to build a guide for implementing these systems effectively in other complex nonlinear systems. Finally, we use the properties of light to transform our photonic system into a computing platform. We investigate experimentally highly nonlinear fibers to replace traditional electronic architecture by leveraging nonlinear pulse propagation.

The results in this thesis compare metaheuristic algorithms for three optical complex systems. We also show the experimental characterizations of a physical artificial neural network. Furthermore, the methods introduced in this thesis demonstrate the potential capabilities when photonics and machine learning techniques are implemented to synergistically support each other.