Dissertation Title: “Machine Learning in Fiber Optics”
Abstract: Recent burgeoning machine learning has revolutionized our ways of looking at the world. Being extraordinarily good at pattern recognition, machine learning has been widely applied to many fields to solve the problems that were previously considered to be extremely challenging. In this dissertation, I demonstrate the applications of machine learning in fiber optics. The dissertation consists of two parts. In the first part, I discuss the use of computer vision in scanning-free fiber-optic imaging. Optical fibers are indispensable tools in endoscopic imaging, due to their miniature sizes and flexibility. Unfortunately, conventional fiber bundles have limited spatial resolutions. In the recent years, multimode fibers (MMFs) have been considered as an alternative, since they have higher mode densities and therefore can carry more information. Imaging through MMFs has been realized by deep supervised learning, where an inverse mapping from the fiber outputs to the objects is learned through numerous pairs of examples. However, the imaging performance suffers from the intrinsic sensitivity of MMFs to perturbations. Altered mode coupling rapidly degrades the image quality. We propose a general framework called the adaptive inverse mapping (AIP) to stabilize the imaging performance through dynamic scattering media, among which MMFs are one example. We show that if the state of the dynamic scattering media is traced closely, the outputs from the media can be used as probes to correct the image reconstruction inverse mapping. As a result, it requires only the outputs and improves the image quality without knowing the sources of perturbations. By using the AIP method, we illustrate the increased robustness of imaging through a MMF under fiber deformations. Despite of the success, the robustness is still rather limited. To further increase the robustness and the image quality, we study imaging through a homemade glass-air Anderson localizing optical fiber (GALOF), where randomness is intentionally introduced in the fiber cross-section. Enabled by the transverse Anderson localization (TAL), the modes in the GALOF are well confined, robust and wavelength-independent. Thanks to these excellent properties, we show that deep unsupervised learning is sufficient for image reconstruction, where the fiber outputs and the object images do not need to be paired for calibration. This greatly simplifies the calibration process and makes imaging through GALOFs flexible. We illustrate robust full-color sub-cellular high-fidelity imaging through the GALOF with deep unsupervised learning. In the second part, I demonstrate the use of machine learning in anti-resonant fiber (ARF) designs. Hollow-core anti-resonant fibers (HC-ARFs) have been widely used for high-power laser delivery, gas lasers and so on. In HC-ARFs, light guidance is based on the capillary structure in the cladding. To achieve desirable fiber propagation properties, various designs of the capillary structure have been proposed in literature. However, the design process so far depends more or less on experience. We propose a reinforcement learning (RL) based method of systematically optimizing the capillary structure to achieve desired properties. As an example, we demonstrate that the RL-enabled HC-ARF designs can achieve an average confinement loss (CL) over a given spectrum with more than one order of magnitude lower than other proposed designs in literature. Moreover, inspired by loss and dispersion spectra of the HC-ARFs, we propose a solid-core anti-resonant fiber (SC-ARF) design for power scaling in fiber lasers at 2 µm. The power levels of the current 2 µm fiber lasers are limited by the modulation instability and soliton formation attributed to the strong anomalous dispersions of fused silica in this wavelength region. Further power upscaling asks for a novel design of an all-solid active fiber that operates in the normal dispersion regime by compensating the material dispersion with the waveguide dispersion, while at the same time achieves a large mode area, low losses, single mode operations and robustness. Taking those goals into consideration, as well as the large design parameter space, including the refractive indices, capillary thickness, etc., it is infeasible to design a SC-ARF by simple parameter sweeps. We utilize a genetic algorithm (GA) to simultaneously optimize the design parameters of the SC-ARF. The proposed SC-HCF exhibits balanced and excellent performances on all the targets, e.g. the smallest dispersion around 2 µm equals to -585.8 ps/nmkm.
BS: 2015, Optical Information Science and Technology, Fudan University
MS: 2020, Optics and Photonics, CREOL, University of Central Florida
Committee in Charge:
Axel Schulzgen, Chair
Rodrigo Amezcua Correa
Jayan Thomas, External
Approved for distribution by Axel Schulzgen, Committee Chair, on June 7th, 2022.
The public is welcome to attend.