Yang, Rui; (2025) Photonic Sampling and Optical Computing Using an Optical Comb and Wavelength-to-Time Mapping. Doctoral thesis (Ph.D), UCL (University College London).

Text Yang_10214923_Thesis.pdf Access restricted to UCL open access staff until 1 October 2026. Download (20MB)

Abstract

Optical fibres, the key component of modern telecommunication, offer ultra-high capacity, low energy loss, and resistance to electromagnetic interference, making them ideal for ultralong-distance communication. Despite these benefits, group velocity dispersion poses a significant challenge, causing optical pulses to spread or broaden as they propagate through the fibre, ultimately limiting the data rate of optical communication systems. Interestingly, this same phenomenon can be leveraged to create advanced optical signal processing techniques. One such example is wavelengthto-time mapping, where varying time delays are introduced across different wavelengths, enabling advanced applications in signal processing. My PhD research primarily focused on investigating advanced photonic information processing techniques based on wavelength-to-time mapping, implemented using single-mode fibres (SMF) and optical frequency combs. This thesis highlights three key techniques developed during this research, including: Firstly, Chapter 3 introduces a tuneable sampling frequency photonic analogueto-digital converter (ADC) as a solution to the fixed sampling frequency limitations of traditional photonic ADCs. This technique leverages a wavelength-to-time mapping established through a 2.4 km SMF and a frequency-agile optical comb, enabling electronically adjustable sampling frequencies of up to 50 GSa/s. To overcome the performance limitations of the sampling pulse train width, Chapter 3 also presents a remote photonic sampling method that employs the compression and subsequent stretching of an optical chirp train. This proof-of-concept experiment successfully demonstrated the technique over a total of 40 km of SMF in laboratory environment, achieving an average improvement of 4.3 dB in the signal-tonoise and distortion ratio compared to the sampling pulses used in the first technique. Finally, a remote 2D convolutional accelerator is introduced in Chapter 4, where a 9.5 km SMF was used for the establishment of wavelength-to-time mapping and the data transmission. The system can simultaneously support the parallel computation of 192 convolutional kernels for deep network, with the computing speed of up to 69 teraoperations per second (TOPS). The system performance was evaluated with a MNIST classification task, achieving an accuracy of 96.2%, closely matching the 96.1% accuracy of to PC-simulated results.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item