Heriot-Watt University
ROS: The Research Output Service. Heriot-Watt University EdinburghNot a member yet
4689 research outputs found
Sort by
Learning priors for scalable computational imaging algorithms, from theory to application in radio astronomy
No full textThis thesis investigates scalable and robust algorithms for image reconstruction, with applications to astronomical imaging. Our approach relies on the versatile framework of plug-and-play (PnP) algorithms, that allows to take advantage of both the robustness of optimisation algorithms through data-fidelity enforcing terms, and of the power of deep neural networks (DNNs) as prior-encoding operators for image restoration tasks. The first section of this thesis deals with theoretical aspects of PnP algorithms from the standpoint of optimisation theory. We investigate conditions in order to ensure the well-definiteness of PnP algorithms and propose two different methods to enforce the associated Lipschitz constraints on the DNNs of interest. As a result, our DNNs behave as resolvents of maximally monotone operators, ensuring a characterisation of the limit point of the associated convergent PnP algorithm. The second section of this thesis applies the proposed methods to radio astronomical imaging, a context where the robustness and scalability of the considered algorithms are paramount. To compensate for the absence of groundtruth data in radio astronomy, we propose a synthetic training dataset with adaptive dynamic range to that serves as a basis for training our DNNs. Results of our PnP algorithms reach similar (when not better) quality than the state-of-the-art, both on simulated and real data, while being significantly faster.James Watt Scholarshi
From crystal to adsorption : new insights into layered double hydroxides derived sorbent materials for carbon capture
Get PDFWith the goal of designing mixed metal oxides (MMOs) that have better CO2 sorption
performance, such as sorption capacities greater than 1.35 mmol/g at the intermediate
temperature range (200 – 400 ˚C), detailed investigation was performance on the MMOs
and their precursor, layered double hydroxides (LDHs). The novelty of work lies in the
crystal chemistry approach, which uses the crystal lattice parameter “a” of precursors
LDHs to obtain the “true” chemical composition of sorbent material. Two systematic
studies based on this approach were conducted to study the effects of different variables on
the CO2 adsorption performance of resultant MMOs, e.g., xcrystal of LDH phase, synthesis
methods and choice of precursors. Synthesis methods was found to be the most important
variable affecting the properties of LDHs and MMOs phase; and co-precipitation method
produces LDH-derived MMOs sorbent with the more desirable CO2 capture performance,
compared to the urea hydrolysis method. The results obtained from these systematic studies
allow the establishment of a new crystal-chemical model for Mg-Al based LDHs that is
fundamentally sound and more accurate when obtaining x from a. Finally, unambiguous
characterization of the equilibrium isotherms and diffusion coefficient of pristine Mg-Al
MMOs was conducted for the intermediate temperature range, using gravimetric and zero
length column method. In the low-pressure region (< 8 bar) and 200 ˚C, the equilibrium
capacities and diffusional coefficient values of pristine MMOs are found comparable to
those promoted with alkali metal salts (AMS). However, in the high-pressure region (8 - 30
bar) and temperature range between 300 – 400 ˚C, the AMS-promoted MMOs shows
almost a two-fold increase in the equilibrium capacities. Overall, the present work set a
reference case for the CO2 adsorption performance of LDH-derived MMOs
An investigation of metalloenzyme biotransformation in a continuous tubular baffled reactor
No full textAbstract unavailable. Please refer to PDF. Restricted access until 31.12.2025
Improved waterflooding performance by low salinity carbonated water injection
No full textAbstract unavailable. Please refer to PDF. Restricted access until 18.01.2026
Climate change and ecotoxicology re-assessing biomarker baselines in light of a changing environment
Get PDFClimate change manifests in the marine environment in several ways. With the present
thesis focusing on increasing sea surface temperature, ocean acidification, and changes in
surface salinity. The temperature of the upper 75 m of the ocean has increased by 0.11°C
per decade since 1971 and pH has decreased at a rate of 0.013 to 0.03 units per decade for
up to 25 years in the ocean surface (these trends are predicted to continue). Additionally,
the surface salinity of the ocean will reflect the increased hydrology cycles of the earth,
with coasts and estuaries likely to experience extreme salinity fluctuations.
Biomarkers might become less sensitive and no longer suitable as indicators of exposure if
climate change causes responses to drift from their current baselines. The purpose of this
thesis was to probe how biomarker responses might be affected by climate change. This
will facilitate future studies being able to distinguish responses caused by non-chemical
confounding stressors from responses to pollutants or anthropogenic activities.
Results indicated that biomarkers of genotoxicity and oxidative stress could be affected by
climatic factors, with interactions between factors affecting statistical significances at
different levels of each factor when combined (e.g., an increased temperature was found to
cause a significant increase in GSH levels when combined with fluctuating salinity, but not
when salinity was static).
The challenge of contextualising results of ecotoxicological studies affected by multiple
stressors in the future has proved to be a very complex task indeed and further knowledge
and resources is needed in this field.James Watt Scholarshi
Harnessing shaped light for enhanced manufacturing using ultrafast laser inscription
Get PDFThe following thesis details the experimental process for the implementation of light beam
shaping into the current standard of ultrafast laser fabrication systems, allowing for the
development of a monolithic microspectrometer device. The implementation of beam
shaping aided in the correction of two limitations in the current writing system: long
manufacturing times and depth-related spherical aberration. Firstly, the addition of a pseudo-Bessel beam allowed for a 12-fold reduction in manufacturing timescales, demonstrated by
chemical etching dicing of a fused silica sample. Moreover, it was shown that polarisation-insensitive selective etching was achieved by tuning the writing laser’s pulse duration.
Secondly, it was shown that with the use of a magnitude-scaled Zernike polynomial phase
mask, the effects of depth related aberration could be counteracted. These experiments
showed that a volume Bragg grating written at a depth of 900 µm produced a low diffraction
efficiency, however, this was increased by 36% when the appropriate phase mask was
applied. This was supported by an electron plasma imaging study which provided further data
without the need for an inscribed object. The knowledge and skills gained during these two
projects produced two published papers and the first steps toward the Czerny-Turner base
monolithic microspectrometer device.EPSRC fundin
Embrace concept drift : a novel solution for online continual learning
Get PDFContinual learning is a critical area of research in machine learning that aims to enable
models to learn new information without forgetting the old knowledge. Online continual learning, in particular, addresses the challenges of learning from a stream of data in
real-world environments where data can be unbounded and heterogeneous. There are two
main problems to be addressed in online continual learning: the first one is catastrophic
forgetting, a phenomenon where the model forgets the previously learned knowledge
while learning new tasks; the second one is concept drift, a situation where the distribution of the data changes over time. These issues can further complicate the learning
process, compared to traditional machine learning.
In this thesis, we propose a general framework for online continual learning that leverages both regularization-based and memory-based methods to mitigate catastrophic forgetting and handle concept drift. Specifically, we introduce a novel concept drift detection
algorithm based on the confidence values of the samples. We present a novel online continual learning paradigm, which utilizes concept drift as a rehearsal signal to improve
performance by consolidating or expanding the memory center. We also apply data condensation approaches to online continual learning in order to perform memory efficient
rehearsal.
Furthermore, we evaluate the accuracy of old tasks and new tasks, comparing with
many benchmark models. We present a novel evaluation metric - Stability and Plasticity
Balance to measure the balance between old and new accuracy.
We evaluate our proposed approach on a new benchmark dataset framework, Continual Online Learning (COnL), which consists of two scenarios of online continual learning: class-incremental learning and instance-incremental learning. In this thesis, the
benchmark dataset framework randomly selects a number of incremental classes from
3 different datasets: TinyImageNet, Germany Traffic Sign and Landmarks. Our primary
results demonstrate that concept drift can be a useful tool in memory rehearsal in the online continual learning setting. Our proposed approaches provide a promising direction
for future research in online continual learning and have the potential to enable models to learn continuously from unbounded and heterogeneous data streams in real-world
environments
Pore scale investigation of surfactant flooding under unsteady-state flow conditions : a combined numerical and experimental study
No full textSurfactant flooding, a chemical Enhanced Oil Recovery (EOR) technique, has been identified as a
potential production protocol for certain oilfields. The standard workflow associated with a
chemical EOR feasibility study include coreflood experiments, which pose several challenges:
they are time-intensive, costly, and often constrained by the availability of suitable core samples.
Pore Network Modelling (PNM) represents an alternative approach that can be used to
complement these coreflood experiments, however, lack of study to fully test the technology for
this purposes. This research consequently aims to develop a validated unsteady-state pore network
modelling simulator that can be used to analyse surfactant flooding in porous media, providing oil
recovery predictions and flow displacement patterns from both pore-scale and core-scale
perspectives.
Laboratory micromodel experiments are also presented to identify the most important surfactant
flooding mechanisms and these observations are then used to construct a detailed numerical PNM
framework. Additionally, coreflood experiments using CT-Scan technology under controlled and
reservoir conditions are described, and these results are analysed and interpreted using PNM
simulations to identify the key parameters affecting surfactant flooding at the laboratory scale.
These include pore structure, frontal velocity, viscosity ratio, wettability (including film flow and
snap off mechanisms), the presence of initial water saturation, and surfactant adsorption. The flow
regime characterising any given coreflood is difficult to predict a priori, as the parameters interact
in complex ways, and so PNM can serve as a useful tool for estimating recovery by surfactant. We
also find that recovery is optimised when surfactant is injected under secondary conditions, whilst
tertiary injection following a waterflood is far less effective.
The newly developed dynamic PNM – validated against experimental data – is shown to be a
robust tool for investigating various surfactant flooding strategies and its potential as a
complementary tool to give insight into coreflood experiments is discussed
Native advertising on the German market : opportunities and constraints for traditional journalism brands
Get PDFThis study examines the possibilities for the traditional German journalism brands to fund
their online editions with native advertising. These advertisements have the appearance
of journalistic articles, are deeply embedded in editorial content, and aim to transfer the
image of the journalistic content to the “camouflaged” advertisement.
In explorative interviews with 22 high caliber industry experts, it could be demonstrated
that at least one third of all digital advertising revenues are generated from native
advertising, predominantly in the form of externally produced programmatic
recommendation advertisements. As early as in the pilot study, it was revealed that these
revenues often have been mistakenly omitted by the industry operatives participating in
national surveys. The interviewees in the main study were therefore encouraged to pay
particular attention to this issue. It could be demonstrated with a high degree of certainty
that recommendation ad revenues have been undervalued in the past.
Whereas publishing houses have been severely impacted from diminishing advertising
budgets, there is an unwavering demand for native recommendation ads. Advertisers are
mainly motivated by the low rates of banner blindness and increasing price
competitiveness because of high user consumption rates. The image transfer mechanism
does not seem to be the main driver for choosing native advertising.
This study was started on the premise that native advertising is mainly defined by branded
journalistic content as described in the academic literature. The research results however
show that they contribute with as little as one tenth to total native advertising revenues
and that recommendation advertisements have become the new cash cows. The third-party providers Outbrain and Taboola have a bilateral oligopoly in this market and their
ad booking platforms have been adopted from advertisers and publishers alike. In
particular, regional and midsized publishing houses have profited from this development
which allows them to gain access to national adverting budgets and increase their ad sales
revenues
Automated laser scanning planning using building information modelling
Get PDFThe current practice in the industry for planning for scanning (P4S) is mainly manual and relies heavily on the surveyor’s judgement to choose scanning locations and
acquisition parameters. The complexity and continuous changes in construction
sites make this task challenging even for experienced surveyors. It is difficult to
ensure that the acquisition campaign fully covers scanning targets while satisfying
quality requirements.
The common approach for Planning for Scanning (P4S) involves formulating the
problem as an optimisation challenge to determine the minimum number of view
points needed to achieve complete coverage of scanning targets while meeting data
quality requirements. By using computational optimisation, scanning operations
can be streamlined to minimise on-site disruptions and data quantity without
compromising the completeness and quality of the data.
Prior P4S techniques have typically relied on a 2D plan view of the scanning
environment as input. However, 2D models lack spatial information that can
impact the quality of P4S output. With the growing availability of 3D Building
Information Modelling (BIM) models, there is a chance to enhance the P4S process
by leveraging the 3D data available.
This thesis introduces an automated P4S algorithm to bridge the knowledge gap,
which takes a 3D BIM model as input and incorporates accuracy, detail, and completeness level requirements. The algorithm also considers scanning overlap and
occlusions. Therefore, the proposed algorithm offers a reliable solution to automate
the P4S problem, resulting in more precise scanning plans that are more likely to
yield point clouds that meet the pre-defined LOA, LOD, and LOC specifications.
Real case studies were used to validate the proposed algorithm by comparing the
outcomes with those achieved by professionals. The findings illustrate the efficacy
of the algorithm, as well as its drawback in terms of computational time. The
research outcome is valuable to both technology-oriented audiences (academics,
researchers, and surveyors) and surveying-oriented audiences (scanning and surveying companies).
The proposed approach guarantees a P4S strategy that achieves the maximum
surface coverage for the target objects while fulfilling the LOA and LOD criteria,
all while suggesting the least number of scanning locations required
Access Repository Dashboard
Do you manage ROS: The Research Output Service. Heriot-Watt University Edinburgh ? Access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard!