Application of Machine Learning to Nonlinear Dynamics in Optical Fibers

Optical pulse propagation in nonlinear optical fibers displays a large variety of complex dynamics. Depending on the regime of propagation, the pulse evolution may be dominated either by coherent dynamics where identical input pulses lead to identical propagation dynamics, or by incoherent dynamics where noise on top of the input pulses can lead to large pulse-to-pulse variation.

The complexity of dynamics in optical fibers can make the analysis and prediction of such systems difficult via conventional techniques. In this thesis, we introduce the use of machine learning techniques to predict and analyse nonlinear pulse propagation in fiber-optic systems. To date, we believe that the potential of machine learning has not been harnessed fully, and this work aims at exploring new approaches and methods in ultrafast photonics based machine learning.

We first demonstrate a rapid and memory-efficient model-free modelling approach to predict the propagation dynamics in the coherent regime of propagation, from the compression of higher-order solitons to broadband supercontinuum generation. Specifically, we demonstrate the application of recurrent neural networks to predict the intensity evolution of pulses without phase information and the application of a simpler feedforward architecture to predict the full-field propagation dynamics. In the second part of the thesis, we introduce techniques for the temporal analysis of ultrafast nonlinear instabilities in fiber-optic systems. Learning from numerical simulations, a neural network is applied to experimental single-shot spectra of noise-seeded modulation instability. The approach is further validated on more complex propagation scenarios: rogue solitons in supercontinuum generation and the analysis of fiber lasers operating in the noise-like pulse regime.

The results presented in this thesis demonstrate one of the first applications of machine learning to nonlinear dynamics and instabilities in optical fibers. Furthermore, the methods introduced in this thesis are very general and can be readily applied to other complex nonlinear systems.