On-chip III-V semiconductor network lasers for neuromorphic computing
File(s)
Author(s)
Dranczewski, Jakub
Type
Thesis
Abstract
Modern Artificial Intelligence workloads are placing a growing amount of strain on digital hardware, increasing costs and energy demands. The field of neuromorphic computing seeks to help alleviate this problem through physical platforms that could be used as accelerators for machine learning tasks. Network lasers are a promising but unexplored platform for this application -- they support a high number of interacting lasing modes through multiple scattering between defined waveguiding paths, and their complexity, non-linearity, sensitivity, and compact size address several challenges facing the field of photonic neuromorphic computing. To enable their use as a computing platform, we need to move them from the current dye-doped electrospun polymer platform to a lithography-defined semiconductor process, which allows for fully designable lasers, and high resiliency to degradation under pumping.
To implement this vision, I first explore the etching of III-V semiconductor layers to create photonic nanodevices. Parameter optimisation allows us to produce high quality structures from InP, our main material of interest, as well as other epitaxial platforms like InGaAs quantum wells. Next, I study the lasing from etched microdisk devices, presenting multiple ways in which lasing properties can be used to inform fabrication choices. With fabrication and lasing performance validated, I present experimental measurements from etched semiconductor network lasers, investigating the physics behind the lasing modes and their coupling. Finally, I show the application of network lasers to neuromorphic computing, first through edge detection, which matches the mode competition dynamics of the laser, and then for image classification, also combining the two schemes to form a photonic convolutional neural network. We obtain high classification accuracies on tasks including MNIST, Fashion MNIST, and BreaKHis 400X, especially in the few-shot regime, and explore the likely physical basis of the neuromorphic performance.
To implement this vision, I first explore the etching of III-V semiconductor layers to create photonic nanodevices. Parameter optimisation allows us to produce high quality structures from InP, our main material of interest, as well as other epitaxial platforms like InGaAs quantum wells. Next, I study the lasing from etched microdisk devices, presenting multiple ways in which lasing properties can be used to inform fabrication choices. With fabrication and lasing performance validated, I present experimental measurements from etched semiconductor network lasers, investigating the physics behind the lasing modes and their coupling. Finally, I show the application of network lasers to neuromorphic computing, first through edge detection, which matches the mode competition dynamics of the laser, and then for image classification, also combining the two schemes to form a photonic convolutional neural network. We obtain high classification accuracies on tasks including MNIST, Fashion MNIST, and BreaKHis 400X, especially in the few-shot regime, and explore the likely physical basis of the neuromorphic performance.
Version
Open Access
Date Issued
2025-06-06
Date Awarded
2026-02-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Sapienza, Riccardo
Moselund, Kirsten
Schmid, Heinz
Sponsor
EU ITN-EID
Grant Number
859841
Publisher Department
Department of Physics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)