Theses - Engineering
Permanent URI for this collection
Browse
Recent Submissions
Item Open Access Assessing the Disruptive Vulnerability of Broadcast News: An Ex-Post Analysis Generating Ex-Ante PrescriptionsLile, StephenThis dissertation endeavors to assess the disruptive vulnerabilities of the broadcast news (BN) industry via a retrospective *ex-post* analysis with the objective of developing forward-looking *ex-ante* prescriptive and performative remedies for BN and the broader journalism industry. Through the lens of Clay Christensen’s disruptive innovation theory (DIT), the analysis begins with demonstrating how the value network, business model, and technology forces of BN, what Christensen refers to as *enablers* of innovation (Christensen, 2009), have become *disablers* of innovation to the detriment of BN and the broader journalism sector. Employing a mixed-methods approach with critical realism and interpretivism as the philosophical underpinnings, the research set out to perform 200 semi-structured interviews (SSI) utilizing a unique and unprecedented research interview method to determine market sentiments about BN and openness to emerging video journalism platforms. SSI questions were informed and cross-referenced by the Pew Research Center, one of the world's leading and most reliable media survey institutions. Qualitative and quantitative data insights were collected and analyzed to determine if there were adverse antecedents suggesting disruptive vulnerability for the industry. The SSI insights were then integrated with learnings from academic literature and proprietary industry sources to generate holistic ex-post insights into the innovation enablers of BN. Foremost of these insights is the breakdown of the BN value network occasioned by five causalities: outmoded institutional logics (legacy knowledge) and losses in credibility, agency, audience, and revenue. BN’s depleted value network has elicited knock-on effects on the other *enablers*, business model and technology, that have become *disablers* of innovation. Social media plays a predominant role in these causalities. Business models have been upended by attention economics machinations where the audience becomes the product sold to advertisers. Technological assaults in the form of AI-driven misinformation and algorithmicizing of user behaviors have led to political polarization and societal unwellness, as affirmed by Pew and this research’s SSIs. This narrative resolves with ex-post insights informing ex-ante prescriptive and performative measures that envision an OTT streaming technology construct to restore BN’s value network, business model, and technology to their healthy, innovation-enabling capacities. The objective is a future-proofed, rehabilitated BN industry powered by cutting-edge digital media logics.Item Open Access Scalable Bayesian Inference in the Era of Deep Learning: From Gaussian Processes to Deep Neural NetworksAntoran, JavierLarge neural networks trained on large datasets have become the dominant paradigm in machine learning. These systems rely on maximum likelihood point estimates of their parameters, precluding them from expressing model uncertainty. This may result in overconfident predictions and it prevents the use of deep learning models for sequential decision making. This thesis develops scalable methods to equip neural networks with model uncertainty. To achieve this, we do not try to fight progress in deep learning but instead borrow ideas from this field to make probabilistic methods more scalable. In particular, we leverage the linearised Laplace approximation to equip pre-trained neural networks with the uncertainty estimates provided by their tangent linear models. This turns the problem of Bayesian inference in neural networks into one of Bayesian inference in conjugate Gaussian-linear models. Alas, the cost of this remains cubic in either the number of network parameters or in the number of observations times output dimensions. By assumption, neither are tractable. We address this intractability by using stochastic gradient descent (SGD)---the workhorse algorithm of deep learning---to perform posterior sampling in linear models and their convex duals: Gaussian processes. With this, we turn back to linearised neural networks, finding the linearised Laplace approximation to present a number of incompatibilities with modern deep learning practices---namely, stochastic optimisation, early stopping and normalisation layers---when used for hyperparameter learning. We resolve these and construct a sample-based EM algorithm for scalable hyperparameter learning with linearised neural networks. We apply the above methods to perform linearised neural network inference with {ResNet-50} (25M parameters) trained on Imagenet (1.2M observations and 1000 output dimensions). To the best of our knowledge, this is the first time Bayesian inference has been performed in this real-world-scaled setting without assuming some degree of independence across network weights. Additionally, we apply our methods to estimate uncertainty for 3d tomographic reconstructions obtained with the deep image prior network, also a first. We conclude by using the linearised deep image prior to adaptively choose sequences of scanning angles that produce higher quality tomographic reconstructions while applying less radiation dosage.Item Embargo Van der Waals Heterostructures and 2D Materials with Native Oxide for Emerging Electronic ApplicationsAlmutairi, AbdulazizThe rapid advancement of Artificial Intelligence (AI) and Machine Learning (ML) in various industries has created a significant energy demand. However, this demand poses a challenge as the world strives to shift towards sustainable energy sources. The current computational paradigms, reliant on complementary metal-oxide-semiconductor (CMOS) transistors, are becoming more inadequate for meeting these emerging demands due to limitations imposed by materials characteristics and device architecture. To address this computational paradigm shift, an innovative material platform is necessary. Two-dimensional (2D) layered materials and their heterostructures offer a promising solution to meet this demand, owing to their unique properties. In addition, oxidation processes of 2D materials, which are termed "morphotaxy", allow for precise control over the material thickness and the fabrication of complex heterostructures, opening the door for advanced computing architectures such as non-von Neumann neuromorphic computing. This thesis explores the critical role of interfaces in the performance and efficiency of devices based on 2D materials van der Waals (vdW) and semiconductor/native oxide heterostructures. The susceptibility of these materials to contamination is highlighted, especially in vdW heterostructures. Elemental analysis of chemical species present at the interfaces of hBN-encapsulated graphene is conducted to determine the origin of contaminants. The findings highlighted the difference between contaminants originating from the heterostructure stacking process and the ones from the device lithography process. Moreover, the impact of such interfaces on carrier transport across two 2D materials and the realisation of phenomena like moiré superlattices in vdW heterostructures based on twisted WSe2/MoS2 hetero-bilayer is discussed. The results showed that inhomogeneity across the heterostructure interface brought by strain and contamination has an impact on the formation of the interlayer exciton (ILX) by reducing the hybridisation of the electronic states between the two 2D layers. In addition, the interface between the semiconductor and oxide in oxidised HfS2 and GaS is probed using complementary characterisation techniques. The results clearly demonstrate the role of the oxidation method on the crystallinity, thickness and uniformity of the produced oxide. This level of control facilitated the production of highly uniform ultrathin oxide layers on top of their corresponding 2D semiconductor materials, which are used for the first demonstration of low-energy resistive switching.Item Open Access Mathematical and computational aspects of solving mixed-domain problems using the finite element methodDean, JosephThis work focuses on mixed-domain problems, where fields defined over different domains are governed by different partial differential equations and may be coupled. Multiphysics problems, where different regions are governed by different (interacting) physical laws, are examples of mixed-domain problems. Mixed-domain problems also arise from the need to employ stable and accurate discretisations tailored to the mathematical nature of the governing equations in each domain. This work develops a new finite element framework for solving mixed-domain problems in FEniCSx that uses efficient, parallel, and scalable algorithms. The framework supports an arbitrary number of domains of possibly different topological dimensions, a range of arbitrarily high-order finite elements, several cell types, and high-order geometry. We demonstrate how solvers for a range of applications can be implemented, including Lagrange multiplier problems, domain decomposition methods, hybridised discontinuous Galerkin methods, and multiphysics problems. Performance results show that the algorithms scale well to thousands of processes. In addition, a hybridised discontinuous Galerkin (HDG) method for the incompressible Stokes and Navier–Stokes equations is generalised to a range of cell types, focusing on preserving a key invariance property of the continuous problem. This invariance property states that any irrotational component of the prescribed force is exactly balanced by the pressure gradient and does not influence the velocity field, and it can be preserved in the discrete problem if the incompressibility constraint is satisfied in a sufficiently strong sense. We derive sufficient conditions to guarantee discretely divergence-free functions are exactly divergence-free and give examples of divergence-free finite elements on meshes containing triangular, quadrilateral, tetrahedral, or hexahedral cells generated by a (possibly non-affine) map from their respective reference cells. We also prove an optimal, pressure-robust error estimate for quadrilateral cells that does not depend on the pressure approximation. The scheme is implemented using the mixed-domain framework and numerical results are provided to support our theoretical analysis. The numerical results also suggest that pressure robustness is preserved on curved cells and that high aspect ratio tensor product cells can be used in boundary layers. A scalable solver is implemented for the statically condensed Stokes system, and performance results show that the scheme is suitable for large-scale problems. A range of multiphysics problems are also considered. We use the mixed-domain framework to employ stable and accurate discretisations in each domain and demonstrate how both monolithic and partitioned coupling schemes can be implemented in a flexible manner. One example focuses on solving the magnetohydrodynamics equations in a domain with both solid and fluid regions. We use a fully coupled (monolithic) scheme that conserves mass exactly and yields a magnetic induction that is exactly solenoidal. Performance results show that the mixed-domain algorithms scale well in parallel.Item Open Access Advances in Meta-Learning, Robustness, and Second-Order Optimisation in Deep LearningOldewage, ElreIn machine learning, we are concerned with developing algorithms that are able to learn, that is, to accumulate knowledge about how to do a task without having been programmed specifically for that purpose. In this thesis, we are concerned with learning from two different perspectives: domains to which we may apply efficient machine learners and ways in which we can improve learning by solving the underlying optimisation problem more efficiently. Machine learning methods are typically very data hungry. Although modern machine learning has been hugely effective in solving real-world problems, these success stories are largely limited to settings where there is an enormous amount of domain-relevant data available. The field of meta-learning aims to develop models with improved sample efficiency by creating models that “learn how to learn”, i.e. models that can adapt rapidly to new tasks when presented with a relatively small amount of examples. In this thesis, we are concerned with amortised meta-learners, which perform task adaptation using hypernetworks to generate a task-adapted model. These learners are very cost efficient, requiring just a single forward pass through the hypernetwork to learn how to perform a new task. We show that these amortised meta-learners can be leveraged in novel ways that extend beyond their typical usage in the few-shot learning setting. We develop a set-based poisoning attack against amortized meta-learners, which allows us to craft colluding sets of inputs that are tailored to fool the system’s learning algorithm when used as training data to adapt to new tasks (i.e. as a support set). Such jointly crafted adversarial inputs can collude to manipulate a classifier, and are especially easy to compute for amortised learners with differentiable adaptation mechanisms. We also employ amortised learners in the field of explainability to perform “dataset debugging”, where we develop a data valuation or sample importance strategy called Meta-LOO that can be used to detect noisy or out-of-distribution data; or to distill a set of examples down to its most useful elements. From our second perspective, machine learning and optimisation are intimately linked; indeed, learning can be formulated as a minimisation problem of the training loss with respect to the model’s parameters — though in practice we also require our algorithms to generalise which is not a concern of optimisation more broadly. The chosen optimisation strategy affects the speed at which algorithms learn and the quality of solutions (i.e. model parameters) found. By studying optimisation, we may improve how well and how quickly our models are able to learn. In this thesis we take a two-pronged approach towards this goal. First, we develop an online hypergradient-based hyperparameter optimisation strategy that improves state of the art by supporting a wide range of hyperparameters, while remaining tractable at scale. Notably, our method supports hyperparameters of the optimisation algorithm such as learning rates and momentum, which similar approaches in the literature do not. Second, we develop a second-order optimisation strategy which is applicable to the non-convex loss landscapes of deep learning. Our algorithm approximates a saddle-free version of the Hessian for which saddle points are repulsive rather than attractive, in a way that scales to deep learning problems.Item Controlled Access Trajectories of Innovation: Measurements, Models & InterventionsHo, Martin; Ho, Martin [0000-0003-2192-6198]Investment in innovation is a crucial determinant of economic development. However, the innovation policy community lacks a robust empirical framework to describe the long-term emergence of technologies. At present, technological progress is often measured quantitatively as the first derivative of incomplete performance metrics as a function of time or interpreted based on qualitative observations at the firm or industry level. There is a longstanding absence of approaches that deterministically trace innovations from their nascent phase – when future applications and potential competitive opportunities may still be unknown – to more applied research, development, and subsequent technological outcomes, such as commercialization. This gap implies the formulation of innovation policy may be blinded by insufficient insights or foresight. The capability to measure historical emergence of technologies would refine the characterization of long-term innovation processes and anticipate how to intervene in these processes. This capability is particularly relevant to “mission-oriented” innovations with predefined technological goals but uncertain innovative intermediaries. This thesis addresses this measurement issue by employing methods from network science and statistics. It posits that various digital records, ranging from patents to clinical trials, and the signals among them, approximate innovation at differing maturities; connecting these records at scale would portray technological emergence more completely. This thesis begins by synthesizing views about innovation “trajectories” from the systems engineering, scientometrics, evolutionary economics, operations research, and political economics literatures. Using these views as a baseline, this thesis derives proof-of-concept methods to trace technological evolution through complex networks from first principles, elucidating sequences of events that represent innovation bottlenecks of vaccines, small molecule drugs, nucleic acid sequencers, large language models, and quantum computers. After using these new methods, this thesis quantifies the position and proximity of documents and funders to long-term innovation events and dynamics behind these emerging technologies. In addition, the new methods are used as a statistical tool to revisit and represent the dynamics of enduring innovation models such as the linear, chain-linked, and S-curve models with empirical data. The most interesting empirical findings arising from the proof-of-concept methods are in terms of the: (i) Order and prevalence of innovation maturities – this thesis observes an overall clear progression of basic research to applied research to development and deployment and that the translation of knowledge is mostly between basic and applied researches and between research and product developments (52% and 36% respectively in this thesis’ dataset); (ii) Division of innovation labor among funding agencies – most large public funding agencies are located 10-27 years from the commercialization of new vaccines while most mission-oriented agencies and pharmaceuticals fund research 2-19 years from commercialization. In addition, most large public funding agencies are similarly involved in early-stage translations, whereas pharmaceuticals and mission-oriented agencies are spread across more diverse innovation maturities, and each leaves different translation footprints. In addition, this thesis develops a (iii) New technique to identify innovations that are rate-limiters to technological emergence. Given the complexity of innovation regularly surpasses the ability of prescriptive models to inform policy decisions holistically, the exploratory methods developed in this thesis present a new and more systematic way to observe and elucidate technological emergence. This work contributes to practice by providing a more objective understanding of innovation “trajectories”, allowing for more informed R&D interventions and policy evaluations.Item Open Access Machine Learning Force Fields for Molecular ChemistryKovács, DávidThe first principles computational modelling of molecular systems is a long-standing pursuit in the scientific community. It has traditionally been tackled by developing approximate solutions to quantum mechanics. Simulations using these electronic structure based methods can be highly accurate but are limited to small system sizes or short time scales. The traditional alternative is force fields that enable fast and accurate simulations by bypassing the treatment of the electrons and describing the system solely in terms of the atomic positions. The emergence of machine learning tools has opened up the opportunity for the development of high accuracy force fields trained directly to reproduce the results of electronic structure calculations. This thesis presents new developments that lead to improved machine learning force fields for molecular chemistry. Firstly, a set of linearly complete basis functions, called ACE, is demonstrated to yield high accuracy custom made force fields for small molecules. By recognising the symmetric tensor structure of these basis functions, the framework is extended to enable the simultaneous description of a large number of chemical elements. Next, multi-ACE is proposed, which provides a unifying theory of most classical and machine learning force fields. Using the design space set out in this theory, a new method, called MACE is created. MACE is shown to provide simple, robust, accurate, and efficient force fields for a wide range of molecular systems. Finally, MACE-OFF23 a new transferable force field for organic molecules is proposed and demonstrated to be capable of accurately describing not only molecules in vacuum but also in the condensed phase.Item Open Access Search Strategies in Holistic Design ProblemsHajnassiri, SaraThe research examines the search process of holistic design in three cases that differ markedly in their complexity, namely graphical design, small household appliance design, and automotive design. Design, as a creative activity, employs a search process because solutions cannot be found by optimisation procedures from input parameters. Holistic designs are characterised by the overall design performance arising from nontrivial interactions among many parameters, so no simple performance function can be described. Structured approaches, such as decomposition by customer needs or by technical systems, have proved to be “insufficient” in addressing such design problems. Using data collected by in-depth interviews, the research compares the search approaches employed in the design process in these distinct industries. Similarities are found in the creative part of the search, where broad solution characterisations and starting points for the search are identified. However, in the experimental exploration of the solution space, the processes differ significantly between cases with different degrees of complexity, even though the search in all cases is built of the two fundamental building blocks of trial-and-error learning and selectionism. For example, the complexity in automotive design leads to the emergence of a highly rugged performance landscape, which requires diligent coordination of teams that search subspaces: Workers searching these subspaces are interdependent, and therefore they need to continually exchange information and adapt to one another. The frameworks resulting from the comparison may help academics and practitioners to better understand design processes from a search perspective. The study may also help managers to decide on the required organisational capabilities and processes for developing products with different degrees of complexity. The thesis concludes by summarising the contributions that were made and suggesting future research opportunities to leverage the work that was initiated and embarked upon with this research project.Item Open Access Data-driven linear interatomic potentialsvan der Oord, CasModelling alloy phase transitions at the atomic scale with first-principles accuracy has been a long-standing research goal. This is becoming increasingly accessible due to the rapid development of data-driven interatomic potential frameworks over the last decade. This thesis presents two novel frameworks, namely 'atomic permutation invariant polynomials' and 'atomic cluster expansion'. Both frameworks are linear models that can be used to create computationally efficient potentials that approximate the potential energy surface for materials and molecules with first-principles accuracy while preserving its underlying symmetries. In addition, this thesis introduces an automated method for building training databases for data-driven interatomic potentials, called 'hyper-active learning'. This method is a pioneering approach that accelerates traditional active learning techniques by biasing (molecular dynamics) simulations towards uncertainty. This novel method is shown to reduce the number of exploratory timesteps required to build a suitable training database by up to an order of magnitude. Finally, a workflow for modelling alloy phase transitions with first-principles accuracy is presented. This workflow involves driving nested sampling simulations using previously mentioned data-driven interatomic potentials to approximate the partition function with first-principles accuracy. From the partition function, free energies and specific heat capacities are then computed, and used to locate first-order phase transitions. These results show good agreement with experimental observations, and are consistent with physically observed phenomena such as the presence of dual-phase micro-structures and precipitate formation in multi-component alloys.Item Embargo Geometry of Creased Shell StructuresMierunalan, SeyonThin Shells are widely used structural elements but are often limited to simple geometries. Special geometric designs can optimise performance through light yet stiffer shell structures. Creasing is one such method that has attracted a lot of traction lately. Many of these structures are underpinned by motion localised within creases, albeit some that exhibit soft modes of deformation from facet bending, leading to non-rigid Origami. These non-rigid Origami structures offer several functional behaviours such as tunable stiffness, deployability, and shape morphing. As such, the potential use of creased shells in a range of applications is intriguing. However, of dominant interest in this thesis is analysing the geometric and mechanical fundamentals of a crease in a shell instead of developing a particular solution. In analysing non-rigid Origami to qualify for practical applications, non-isometric mechanical models are crucial to determine the shape and mechanics of the resulting structure. The simplification of finite element modelling proposed in this thesis offers a solution that can capture the geometric intricacies and mechanics of non-rigid Origami which are not captured by simple isometric models, with computational efficiency. This validated FEM has been used to explore the geometry and mechanics of simple but interesting non-rigid Origami structures throughout the thesis, and compared against simple analytical models. A straight crease embedded in a thin metal sheet serves as a starting point for exploring non-rigid Origami. This is approached by classifying the problems based on the location of the crease ends, and geometric constraints. Simple energy arguments along with experimental and finite element models are utilised to capture the underlying mechanics. It is found that the geometry of the creased sheets with a single straight crease depends on the location of the crease end, the extent of the sheet, the thickness of the sheet, and the degree of deformation or the fold-angle. Following this, a similar approach is employed to investigate curved creases. Unlike straight creases, the creasing method influences the deformation of the sheet with a curved crease, the effect of which is explored using differential geometry and simple plate buckling analogy where a good agreement with finite element models are observed. This reveals that a curved crease introduces mechanical frustration in one way or the other. The finite element solutions capture the many nuances in geometry of the curved creased sheets that analytical models with isometric assumption overlook. Finally, a posteriori from the two single crease studies are utilised to explore several cases of intersecting creases and understand their geometry and force-displacement mechanics. A parametric study revealed the geometric and material parameters that affect the bistability of mulitply-creased sheets, which can then be fine tuned to suit a wide range of applications. Throughout this thesis, the intention has been to pursue simple analytical methods supported by informal/formal observations obtained from small-scale physical experiments, to derive generalised insights and relationships. These will serve as a building block for applications of non-rigid Origami using straight/curved creases. Insights into the straight and curved creases and associated compliance and bistability from this study can be leveraged by engineers for a range of novel technologies from nano-scale air vehicles, developable mechanisms, switches, actuators to macro-scale satellite technologies.Item Open Access Polyhedral Computation for Differential System Analysis and ControlKousoulidis, Dimitris; Kousoulidis, Dimitris [0000-0002-1508-2403]In this thesis we investigate the use of polyhedra in the analysis and design of dynamical systems. The main motivation behind the use of polyhedra in this context is that they can, in principle, provide arbitrarily tight conditions on stability, monotonicity, and some system gains for a large class of systems. However, finding a suitable polyhedron is a difficult problem. This is to a large extent inevitable, since many of the above problems are known to be computationally intractable. Despite this, the conditions that a polyhedron must satisfy in the above problems have a strong geometric intuition and a fundamental connection to Linear Programming (LP), allowing for the development of effective and sound heuristics. These can be very valuable because they allow us to better leverage all computational power available and because for many practical scenarios, especially those involving design, tight results might not be necessary but any improvements over existing relaxations are still beneficial. The main contribution of this thesis is the development, presentation, and evaluation of such heuristics for variations of the aforementioned problems. A central idea is the use of LP not only to verify conditions for a given polyhedron, but also to iteratively refine a candidate polyhedron through a local optimisation procedure. This allows for a fine-tuned trade-off between conservativeness and computational tractability and can be used for both analysis and design. For each of the problems considered we also include numerical case studies that demonstrate the effectiveness of this idea in practice. However, more work is necessary to establish theoretical guarantees about the performance and convergence of this approach. We also provide a unified exposition on polyhedra with a focus on computational considerations and the differences between their two representations, including a novel characterisation of the subdifferential of polyhedral functions in one of the representations that leads to novel dissipativity conditions for bounding the L1 gain of systems. Differential analysis is used throughout to link the conditions on the polyhedra to the resulting system behaviour. We hope that this research broadens the applicability of polyhedral computation in systems and control theory and opens a promising avenue for future research.Item Open Access Development and Optimisation of an Optofluidic Evanescent Field Nano Tweezer System for Trapping Nanometre Crystals for Synchrotron X-Ray Diffraction ExperimentsDiaz, AlexandreX-ray crystallography (the analysis of the diffraction pattern of a high-power X-ray) has become the preferred method for investigating complex molecular structures. Research has continued to improve both the hardware and software elements of this methodology; refining critical factors such as detectors, beam generation and post modelling computation to keep increasing performance. One overlooked, yet critical factor is increasingly impacting the developmental progress of this technology, sample loading. This is currently achieved via expensive, electromechanical and robotics systems, but technological advancement can increase measurement accuracy by improving positional accuracy, maintaining the crystal integrity by keeping them in their native solution and reduce experimentation times through faster loading; factors which are all currently limiting the minimum sample crystal size to >5 µm. A by product will be a reduction in investigation costs by reducing beamline set-up and testing time, potentially opening the technique to other research fields, e.g. biomedical, archaeological or engineering. This work focusses on a specific sub-section of the loading challenge, introducing a method which could achieve reliable sample loading of crystals below 5 µm, thus opening up the low micro and nano crystallography frontier. The solution proposed in this thesis implements a novel sample loading method, an optofluidic system which combines the advantages of both evanescent field optical tweezing and microfluidics. A low cost commercial Nanotweezer system and optofluidic chip technology were assessed via interferometry and SEM microscopy to determine the microfluidic and waveguide architecture. It was shown that across several chip batches the waveguide dimensions remained similar but with notable structural variations due to imperfect production and poor post manufacturing modifications, highlighting the need for an alternative chip design and manufacturing process. Factors affecting the sample generation and stabilization, crystallisation and microfluidic parameters were also investigated through a series of tweezing trials on latex nanospheres and lysosome crystals. Results indicated that an operational temperature of 2°C and a horizontal chip orientation (0° being optimal, but effective up to 60°) were the most critical factors. COMSOL modelling highlighted that both the evanescent field form and intensity could be tailored to the crystal size and shape via the addition of either “doughnut” or “bullseye” nano plasmonic antennas on the surface of the optofluidic chip waveguides. When combined with a 1-D PhC resonator array they formed a hybrid waveguide that increased the electric field intensity from 1.10 × 107 V/m to 1.70 × 107 V/m. A prototype manufacturing route for the refined architecture was evaluated. A first batch of upgraded optofluidics chips was upgraded using focused ion beam assisted gas deposition to generate the platinum plasmonic antennas. This method however, proved unsuccessful so the remaining batch was upgraded using e-beam lithography to generate gold plasmonic antennas. While satisfactory launching of the newly upgraded chips was not achieved due to technical limitations, characterisation testing of the default waveguide demonstrated microscale (2 µm) and sub-micron (0.8 µm) tweezing, and that such performance could theoretically be enhanced with the addition of the plasmonic antenna structures.Item Open Access Learning Monocular Cues in 3D ReconstructionBae, Gwangbin; Bae, Gwangbin [0000-0003-4189-3493]3D reconstruction is a fundamental problem in computer vision. While a wide range of methods has been proposed within the community, most of them have fixed input and output modalities, limiting their usefulness. In an attempt to build bridges between different 3D reconstruction methods, we focus on the most widely available input element -- *a single image*. While 3D reconstruction from a single 2D image is an *ill-posed* problem, monocular cues -- e.g. texture gradients and vanishing points -- allow us to build a plausible 3D reconstruction of the scene. The goal of this thesis is to propose new techniques to learn monocular cues and demonstrate how they can be used in various 3D reconstruction tasks. We take a data-driven approach and learn monocular cues by training deep neural networks to predict pixel-wise surface normal and depth. For surface normal estimation, we propose a new parameterisation for the surface normal probability distribution and use it to estimate the aleatoric uncertainty associated with the prediction. We also introduce an uncertainty-guided training scheme to improve the performance on small structures and near object boundaries. Surface normals provide useful constraints on how the depth should change around each pixel. By using surface normals to propagate depths between pixels, we demonstrate how depth refinement and upsampling can be formulated as a classification of choosing the neighbouring pixel to propagate from. We then address three challenging 3D reconstruction tasks to demonstrate the usefulness of the learned monocular cues. The first is multi-view depth estimation, where we use single-view depth probability distribution to improve the efficiency of depth candidate sampling and enforce the multi-view depth to be consistent with the single-view predictions. We also propose an iterative multi-view stereo framework where the per-pixel depth distributions are updated via sparse multi-view matching. We then address human foot reconstruction and CAD model alignment to show how monocular cues can be exploited in prior-based object reconstruction. The shape of the human foot is parameterised by a generative model while the CAD model shape is known *a priori*. We substantially improve the accuracy for both tasks by encouraging the rendered shape to be consistent with the single-view depth and normal predictions.Item Open Access On Aerothermal Optimization of Low-Pressure Steam Turbine Exhaust SystemCao, JiajunThis thesis addresses two challenges in the aerothermal optimization of Low-Pressure Steam Turbine Exhaust System (LPES). The first one is the high computational cost due to the complexity of LPES geometry. To make things worse, designers have to consider extra cost caused by change of constraints, multi-objective and multi-disciplinary optimization. The second is the vulnerability of optimization due to the lack of comprehensive validation of numerical simulation. Sparse experimental data from LPES rig can only give limited and sometimes misleading information for validation. To reduce the computational cost, a commonly used way is to build a surrogate model. However, manual parametrization of high-dimensional geometries like LPES is unreliable. Thus, a Non-Parametric Surrogate Model (NPSM) is developed, which directly builds a mapping relationship between surface mesh and two-dimensional distribution of fluid variables. It can select sensitive geometric features from surface meshes by Graph Neural Networks (GNNs) encoder according to the back-propagated prediction error, which reduces uncertainties caused by manual parameterization and gains the ability to process designs defined by different geometry generation methods. Based on NPSM, a non-parametric sensitivity analysis is conducted, which can calculate the distribution of sensitivity on the surface meshes. It can help users to identify important geometric features and redistribute the control points of geometry. Furthermore, a design classifier is built to detect predictable designs for NPSM, thereby preventing compromises to robustness of optimization. To enhance the robustness of optimization, validation of numerical simulation by exper iment is essential, but the sparsity of experimental data due to the large volume of LPES prevents a comprehensive comparison. This thesis demonstrates a Physical-Informed Neural Networks (PINNs)-based method to reconstruct sparse data, which has much better perfor mance than interpolation. In addition, it can be used to detect anomalies, which prevents data contamination due to mistakes in experiment. A Non-Uniform Rational B-spline (NURBS)-based optimization algorithm is also pre sented in this thesis. It generates surface meshes for NPSM and volume meshes for CFD solver based on control points given by optimizer. The conversion process is achieved by the evaluation of NURBS surfaces and parabolic mesh generator, which provides more degrees of freedom and keep mesh generation robust.Item Embargo High Bandwidth Rogowski Coils, Commutation Loop Stray Inductance, and the Si IGBT and the SiC MOSFET Based Hybrid ConceptZhang, TianqiWide bandgap (WBG) semiconductor power devices are attracting increasing attention due to their superior performance across various applications. For accurate measurement of these devices, a current sensor with a minimum bandwidth of 70 MHz is required. This thesis thoroughly explores the design and analysis of both toroidal and solenoidal printed circuit board (PCB) Rogowski coils. The associated parameters of these Rogowski coils are meticulously extracted using ANSYS Q3D Extractor and subsequently measured with the Tektronix TTR500 vector network analyser (VNA). Following this, the thesis presents the design of an op-amp-based integrator, with parameters imported into LTspice for simulation. The bandwidths of the Rogowski coil-integrator assemblies are measured using the VNA, which reveals that the bandwidth of a 10-turn solenoidal PCB Rogowski coil impressively exceeds 300 MHz. Moreover, a comparative analysis of handmade coils and PCB coils is conducted. A relay-based edge generator, designed to facilitate time-domain testing of Rogowski coils, is capable of producing a voltage rise from 10% to 90% in just approximately 6 ns. The Rogowski coil measurement closely correlates with the onboard current measurement, thus confirming the validity and effectiveness of the design approaches presented within this thesis. This thesis also provides a comprehensive introduction to the fundamental aspects of the metal-oxide-semiconductor field-effect transistor (MOSFET) and the insulated-gate bipolar transistor (IGBT), deepening the reader’s understanding of these essential components in power electronics. It then focuses on the current commutation loop inductance, underscoring the importance of minimising this parameter to optimise the design and efficiency of converters and modules, especially when employing WBG power devices in high-speed applications. The thesis introduces various strategies to mitigate loop stray inductance, emphasising the use of compact PCB layouts and laminated bus bars. Additionally, it presents an experimental methodology for extracting loop inductance values, illustrated with a practical example that demonstrates the extraction process. To fully capitalise on the advantages of the Si IGBT’s low conduction loss and the silicon carbide (SiC) MOSFET’s low switching loss, a hybrid parallel connection of a Si IGBT and a SiC MOSFET is proposed. Six gate signal control strategies are introduced and evaluated. The total power loss and device junction temperatures across various load currents and switching frequencies are simulated, investigated and analysed using LTspice and PLECS. For light loads, the SiC MOSFET operates independently. In the case of heavy loads, the Si IGBT takes charge of conduction, while the SiC MOSFET serves as a current bypass to assist in IGBT switching. Simulation results indicate that the hybrid parallel connection can significantly reduce both the total loss and thermal requirements. Moreover, the hybrid parallel connection can also minimise the impact of the freewheeling diode’s (FWD) reverse recovery during the IGBT’s turn-on phase.Item Open Access Active and Semi-Supervised Learning for Speech RecognitionKreyssig, FlorianRecent years have seen significant advances in speech recognition technology, which can largely be attributed to the combination of the rise in deep learning in speech recognition and an increase in computing power. The increase in computing power enabled the training of models on ever-expanding data sets, and deep learning allowed for the better exploitation of these large data sets. For commercial products, training on multiple thousands of hours of transcribed audio is common practice. However, the manual transcription of audio comes with a significant cost, and the development of high-performance systems is typically limited to commercially viable tasks and languages. To promote the use of speech recognition technology across different languages and make it more accessible, it is crucial to minimise the amount of transcribed audio required for training. This thesis addresses this issue by exploring various approaches to reduce the reliance on transcribed data in training automatic speech recognition systems through novel methods for active learning and for semi-supervised learning. For active learning, this thesis proposes a method based on a Bayesian framework termed NBest-BALD. NBest-BALD is based on Bayesian Active Learning by Disagreement (BALD). NBest-BALD selects utterances based on the mutual information between the prediction for the utterance and the model parameters, i.e. I[θ, w|Dl, Xi]. Monte-Carlo Dropout is used to approximate sampling from the posterior of the model parameters and an N-Best list is used to approximate the entropy over the hypothesis space. Experiments on English conversational telephony speech showed that NBest-BALD outperforms random sampling and prior active learning methods that use confidence scores or the NBest-Entropy as the informativeness measure. NBest-BALD increases the absolute Word Error Rate (WER) reduction obtained from selecting more data by up to 14% as compared to random selection. Furthermore, a novel method for encouraging representativeness in active data selection for speech recognition was developed. The method first builds a histogram over the lengths of the utterances. In order to select an utterance, a word length is sampled from the histogram, and the utterance with the highest informativeness within the corresponding histogram bin is chosen. This ensures that the selected data set has a similar distribution of utterance lengths to the overall data set. For mini-batch acquisition in active learning on English conversational telephony speech, the method significantly improves the performance of active learning for the first batch. The histogram-based sampling increases the absolute WER reduction obtained from selecting more data by up to 57% as compared to random selection and by up to 50% as compared to an approach using informativeness alone. A further contribution to active learning in speech recognition was the definition of a cost function, which takes into account the sequential nature of conversations and meetings. The level of granularity at which data should be selected given the new cost function was examined. Selecting data on the utterance-level, as fixed-length chunks of consecutive utterances, as variable-length chunks of consecutive utterances and on the side-level were examined. The cost function combines a Real-Time Factor (RTF) for the utterance length (in seconds) with an overhead for each utterance (t1) and an overhead for a chunk of consecutive utterances (t2). The overhead t2 affects the utterance-level selection method (which previous methods in the literature rely on) the most, and this level of granularity yielded the worst speech recognition performance. This result showed that it is crucial to focus on methods for selection that can take a better cost function into account. For semi-supervised learning, the novel algorithm cosine-distance virtual adversarial training (CD-VAT) was developed. Whilst not directed at speech recognition, this technique was inspired by initial work towards using consistency-regularisation for speech recognition. CD-VAT allows for semi-supervised training of speaker-discriminative acoustic embeddings without the requirement that the set of speakers is the same for the labelled and the unlabelled data. CD-VAT is a form of consistency-regularisation where the supervised training loss is interpolated with an unsupervised loss. This loss is the CD-VAT-loss, which smoothes the model’s embeddings with respect to the input as measured by the cosine-distance between an embedding with and without adversarial noise. For a large-scale speaker verification task, it was shown that CD-VAT recovers 32.5% of the Equal Error Rate (EER) improvement that would be obtained when all speaker labels are available for the unlabelled data. For semi-supervised learning for speech recognition, this thesis proposes two methods to improve the input tokenisation that is used to derive the training targets that are used in masked-prediction pre-training; a form of self-supervised learning. The first method is biased self-supervised learning. Instead of clustering the embeddings of a model trained using unsupervised training, it clusters the embeddings of a model that was finetuned for a small number of updates. The finetuning is performed on the small amount of supervised data that is available in any semi-supervised learning scenario. This finetuning ensures that the self-supervised learning task is specialised towards the task for which the model is supposed to be used. Experiments on English read speech showed that biased self-supervised learning can reduce the WER by up to 24% over the unbiased baseline. The second method replaces the K-Means clustering algorithm that was previously used to tokenise the input with a Hidden Markov Model (HMM). After training, the tokenisation of the input is performed using the Viterbi algorithm. The result is a tokenisation algorithm that takes the sequential nature of the data into account and can temporally smooth the tokenisation. On the same English read speech task, the HMM-based tokenisation reduces the WER by up to 6% as compared to the tokenisation that uses K-Means.Item Open Access Atomically thin microwave biosensorsGubeljak, Patrik; Gubeljak, Patrik [0000-0001-6955-419X]This thesis described the development of a new type of biosensor based on the integration of atomically thin graphene in a broadband radio-frequency (RF) coplanar waveguide (CPW). To combine both electrochemical and dielectric sensing concepts, a graphene channel is inserted into the CPW, which can then be functionalised using standard chemical processes developed for graphene direct-current (DC) sensors. The new RF sensors shown in this work inherit the strong response of graphene at low analyte concentrations, while enabling RF methods at concentrations not possible before. This is shown by limit of the detection for glucose, based on RF measurements, several orders of magnitude lower than that previously reported in literature, enabling measurements of glucose at levels found in human sweat. The combined effect of the two sensing responses can be seen as the high sensitivity at low concentrations, 7.30 dB/(mg/L), significantly higher than metallic state-of-the-art RF sensors, which gradually decreases as the graphene saturates and the response at higher concentrations aligns with that of previously reported metal sensors. This shows the extended operating range of the new sensors compared to the literature state-of-the-art. Similarly, for DNA strands, the graphene CPW sensors was able to distinguish between different types of DNA (perfect match, single mutation, and complete mismatch) at the limit of detection, 1 aΜ, improving the concentration response by orders of magnitude compared to existing RF DNA sensors and allowing for distinction of the types at lower concentration than those in contemporary DC graphene sensors. The additional effect of tunable graphene conductivity independently of the analyte, is the creation of multidimensional datasets for each *S*-parameter component by combining the frequency and potential responses. The resultant sensor response surfaces contain more information that can be used to determine the concentration and type of the analyte. In particular for DNA, the multidimensional approach, by considering the frequency and biasing condition for the largest joint changes in the parameters allowed for direct classification of the three DNA analyte types based on the sign of the changes in the parameters ($ℜ/ℑS_{ij}$ or $mag/∠S_{ij}$). Because of the multidimensional nature of the dataset, machine learning algorithms could be applied to extract features used for determining concentrations based on the raw measurements. For glucose the concentration was predicted with larger than 98% confidence, while the more feature-rich surfaces of the DNA response allowed the determination of the analyte type at 1 aΜ even with a simulated signal-to-noise ratio of 10 dB, proving the utility and resilience of the developed devices and measurement approach. The above approaches were only enabled by the new type of sensor combining both graphenes electronic properties and RF sensing concepts in a single device.Item Open Access Humanitarian Sheltering: Analysing Global Structures of AidGeorge, Jennifer; George, Jennifer [0000-0002-0580-7386]The provision of shelter is an integral part of humanitarian response, in aiding communities affected by crises, in post-disaster, post-conflict and complex situations. Indeed, prior research has identified the wider impacts that shelter can have in these situations, across health, livelihoods, economic stimulation, education, food and nutrition, and reducing vulnerability. However, there is still a lack of understanding of the processes involved in shelter programming, the key decisions directing shelter response, and the influence that different stakeholders hold over those decisions. This thesis analyses the global shelter sector though a systems-thinking approach, including the structures which affect behaviour in this system, the relationship between different actors, and the relationship between decisions taken over time. It identifies key decision makers and decision moments which directly and indirectly influence outputs in shelter projects, analysing controls on decision-making and the complexity of humanitarian governance in shelter projects. This is achieved through expert interviews, analysis of historic cases of shelter and current guidance, and participant observation. This research reveals that community involvement in decision-making is often a very constrained exercise, despite repeated rhetoric over its necessity for project success. It also illustrates the top-down power dynamics that exist in decision-making, oftentimes hidden behind the supposed technocratic focus of the shelter and settlements sector. This includes influence over projects by donors, governments, the humanitarian system, private sector, and public opinion. It examines perceived constraints in-depth, including donor policies and funding timelines, political priorities of national governments, humanitarian mandates and priorities, private sector partnerships, iconography of shelter, and the role of affected populations themselves. This thesis will show that humanitarian shelter should be re-defined at a policy level as ‘an enabled process to facilitate a living environment with crisis-affected communities and individuals to meet their current and future needs, whilst also having due consideration for the needs of the host communities and environment’. This is required to shift perceptions of shelter across actors who are traditionally outside of the shelter sector and incorporate the learning in shelter and settlements that has occurred over the last forty years.Item Embargo Conductive PEDOT:PSS fibres for modelling and assessing oligodendrocyte ensheathmentLiu, RuishanOligodendrocytes are recognised for their capacity to ensheath neuronal axons with tightly packed, multi-layered cell membranes, forming the myelin sheath. This myelin sheath plays a pivotal role as an insulating layer surrounding neuronal axons. It effectively segregates the conductive environments inside and outside neuronal axons, thereby preventing ion leakage and minimising signal loss. The loss of the myelin sheath can lead to physical and mental health issues, including disabilities and depression. Consequently, there is a pressing need for in-vitro models to investigate oligodendrocyte ensheathment behaviour. Current models often employ engineered nano- or micro-fibres to simulate neuronal axons but frequently lack electrical conductivity. To address this gap, this thesis introduces novel conductive fibres made from Poly(3,4-ethylenedioxythiophene):Poly(Styrene Sulfonate) (PEDOT:PSS). These biocompatible fibres possess electrical conductivity and mechanical stiffness, mimicking the diameter of axons while facilitating electrical stimulation and real-time impedance measurements. Examination of Scanning Electron Microscopy (SEM) images reveals that oligodendrocytes adhere to these fibres, extend their cell membranes, and ensheath the fibres. Immunofluorescence images further indicate that the oligodendrocyte expresses Myelin Basic Protein (MBP), which is a characteristic of myelinating oligodendrocytes and has the ability to compact multiple cell membrane layers. Additionally, the author has derived an equivalent circuit to interpret device-level impedance results.Item Controlled Access From Pine Cones to Minimal Surfaces: The Geometry and Mechanics of Morphing BilayersSalsby, BarneyUbiquitous to nature, is an inherent ability to program form to fulfill function. In thin biological structures, such as leaves and plants, growth coupled with their slenderness, allows for a wealth of complex geometries to emerge. Underpinning this sophisticated phenomena is a problem of mechanics, in which growth, or more specifically volume changes, impart inhomogeneous strain profiles which frustrate the structure and trigger deformations by way of buckling or wrinkling. Inspired by this, researchers have mathematically and physically characterised these phenomena, enabling 2D sheets to be programmed into 3D shapes and reconfigure between their different stable forms. This would have wide ranging applications in a variety of contexts such as robotics for instance, where smart actuators and mimicking live tissue is needed, or for deployable structures in aeronautics. A subset of this type of growth is that of bilayers, where the Uniform Curvature model has enabled researchers to successfully investigate problems of buckling and multistability. However, owing to the free edge, this model fails to capture higher order effects pertaining to a boundary layer phenomenon, in which moments must dissipate, and hence geometry vary beyond the quadratic terms near the boundary. Traditionally, researchers have neglected their effect in view of the bulk behaviour. However, the resulting linearly scaled Gauss for the stretching energy does not distinguish between planform geometries, and given recent findings concerning the influence of edge effects on preferred bending direction, the validity of the Uniform Curvature model has been put into question. By introducing a ’fictitious’ edge moment rotation, the energy function is reduced commensurately with the dissipation that occurs within this boundary layer. These reduction terms are a function of the curvatures and we obtain a system of algebraic equations. By consideration of the neutrally stable shell, we observe the altering in preferred bending configuration and stability properties due to planform geometry effects, which we validate by way of a physical prototype. By introducing a non-linear scaling of Gauss for the stretching energy, we further investi- gate the cessation of multistability into monostability as aspect ratio is varied. By further coupling this non-linear scaling term with edge effects, we uncover a novel tristable structure and demonstrate how it can be straightforwardly fabricated in a table-top experiment. By use of soft elastomers, we investigate the one- and two-dimensional de-localisation of the boundary layer, noting a minimal surface for opposite-sense prestressing for two-dimensional de-localisation. This dissertation thus provides an insight into the role edge effects play on bilayers in the context of disparate planform geometries, which we further combine with a non- linear variation of the stretching energy, to see how the coupling of these accounts for the multistable properties and geometry as aspect ratio is varied. Beyond the insights expounded, the approach extends the Uniform Curvature model for the study, design and fabrication of morphing bilayers and their subsequent applications.