Postgraduate Theses

Enhancing Security and Privacy in Cryptocurrency Exchanges

2026-01-01T00:00:00Z

Enhancing Security and Privacy in Cryptocurrency Exchanges Chen, Quanhao Centralized exchanges (CEXs) dominate cryptocurrency markets due to liquidity and low latency, but their opaque internal ledgers create custodial risk. Meanwhile, privacy concerns motivate private exchanges that blind the platform to user balances and order flow. Recent work such as Pisces [1] explores private and compliable exchanges, but one critical piece in compliance, public verifiable full solvency, remains unresolved. When liabilities are hidden from the platform, the platform cannot construct plaintext-based commitments and cannot be trusted to disclose complete liability sets at audit time, creating a fundamental privacy–solvency conflict. We address this conflict by designing two systems that enforce both solvency and platform-side privacy: * **Audit-then-Check Private and Solvent Exchange System:** Uses an RSA accumulator to provide constant-size membership witnesses; users verify inclusion after the auditor publishes an audit snapshot, and omission yields publicly verifiable evidence. * **Certify-then-Audit Private and Solvent Exchange System:** Eliminates user participation by introducing trusted hardware that certifies each transaction acceptance via a monotonic counter log, enabling the auditor to verify completeness without learning transaction contents. We provide rigorous security analysis that formally establish privacy against a malicious platform and solvency soundness against a malicious user-platform coalition. We implement both schemes and evaluate performance against the state-of-the-art baselines. Our prototype achieves average per-procedure computation under 35 ms, communication bounded by 14 KB per operation. For solvency verification, our online prover time remains nearly constant across user scales, achieving a 41.6× speedup over the state of the art and a 268.6× speedup over deployed baselines at $N = 2^{14}$ users, demonstrating that frequent auditing remains feasible at scale.

An Analysis of the Australian Patient-Sharing Provider Networks on Healthcare Provider Cost Performance Using Administrative Claim Data

2026-01-01T00:00:00Z

An Analysis of the Australian Patient-Sharing Provider Networks on Healthcare Provider Cost Performance Using Administrative Claim Data Karim, Shakir Mohammad Shafayat Patient-sharing provider networks (PSPNs), which are based on how providers collaborate to care for patients, provide a good indication of how provider interactions affect the system's operation. Although more people are interested in network-based solutions, there is insufficient real-world evidence linking the PSPN structure to provider cost performance, particularly in the Australian healthcare system. This research examines the correlation between PSPNs and provider cost performance in the Macarthur Area, the Parramatta–Hills, and the Sydney metropolitan area in New South Wales, Australia. The study leverages comprehensive administrative claim data to create PSPNs based on patient admission experiences. It utilises social network analysis to assess the structural roles of providers. Degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, and clustering coefficient are some network metrics that show how providers work together. Providers' cost performance is determined by specific administrative cost metrics that enable systematic comparisons across network sites and geographic regions. This research integrates network modelling with non-parametric statistical methodologies suited to the distributional properties of administrative claim data. Spearman's rank correlation analysis is utilised to examine monotonic relationships between network centrality measures and provider cost performance. Stratified studies compare very central providers with peripheral providers to evaluate structural variability within networks. These methods yield robust conclusions that are not dependent on strict parametric assumptions. The data yield diverse outcomes; some indicate statistically significant correlations between provider network location and cost performance, whereas others do not. The results show that integration in patient-sharing provider networks affects how care is coordinated,and how cost-effective treatment is. Includes publication

Lightweight Deep Learning Architecture for Image Computing

2026-01-01T00:00:00Z

Lightweight Deep Learning Architecture for Image Computing Guan, Jingwen As deep learning has evolved from early convolutional neural networks (CNNs) to deep residual networks, attention mechanisms, Transformer architectures, and even large-scale pre-trained models, continuous breakthroughs in model performance have been accompanied by exponential growth in parameter size and computational overhead. This poses significant challenges for deployment in resource-constrained scenarios such as mobile devices and edge computing. Therefore, the current research focus has gradually shifted from solely pursuing extreme performance to the synergistic optimisation of performance and efficiency. To achieve a lightweight deep learning structure, existing research has explored both network architecture design and training paradigms. At the network level, techniques such as multi-scale feature fusion, attention mechanisms, and other structural enhancements are employed to enhance feature capture ability. However, these methods often rely on additional modules and parameters, which inevitably increase model complexity and further complicate the challenging process of feature learning and optimisation. At the training level, diverse forms of loss design and training philosophies (such as contrastive learning, self-supervised learning, and distillation) have been employed to enhance the generality and robustness of learned features. Yet, whether gradient signals during backpropagation that modify feature encoding are susceptible to interference from noise, bias, and other factors remains inadequately characterised theoretically and empirically. These issues indicate that achieving lightweight feature encoding requires not only structural changes, but also a system-level, cooperative design. This thesis aims to explore the enhancement of feature encoding capabilities across three levels: the training level, the convolutional layer level, and the network level. Includes publication

Enhancing Medical Vision-Language Foundation Models in Tumour Malignancy Recognition via Pre-trained Language Models

2026-01-01T00:00:00Z

Enhancing Medical Vision-Language Foundation Models in Tumour Malignancy Recognition via Pre-trained Language Models Wang, Xiao Vision-Language Foundation Models (VLFMs) offer strong potential for zero-shot learning in histopathology, where models must recognise rare or unseen disease subtypes with limited annotation. However, pathology VLFMs often rely on coarse textual descriptions that miss fine-grained visual cues, causing weak image-text alignment, noisy retrieval, and reduced classification accuracy. This thesis proposes Retrieval-based De-noising Causal Language Modelling (RDCLM), a framework for improving zero-shot tumour malignancy recognition by refining noisy retrieval outputs from pathology VLFMs. RDCLM builds a pathology-specific knowledge base of discriminative benign and malignant tumour descriptions using a large language model. For each query histopathology image, a pathology VLFM retrieves candidate descriptions from this knowledge base, and a frozen pre-trained language model integrates the retrieved text with projected visual features to suppress irrelevant content and retain malignancy-relevant evidence. To improve robustness, two retrieval augmentation strategies are introduced: Retrieval Negatives Replacement (RNR) and Description-wise Shuffling (DS). A Multi-Branch Diverse-Dimension Projection architecture with an auxiliary Inter- and Intra-Branch Min-Max Mutual Information Optimisation objective is also proposed to encourage diverse and informative cross-modal feature learning. Precision Reward Loss is further evaluated as an auxiliary refinement for cleaner de-noised descriptions. Experiments on five histopathology cancer datasets show that RDCLM improves both zero-shot image-text retrieval and image classification compared with state-of-the-art CLIP-based, retrieval-based, and unimodal baselines. The results demonstrate that retrieval de-noising with pre-trained language models can strengthen semantic alignment and improve VLFM-based malignancy recognition in histopathology. Includes publication

Synergistic Embodiment: A Framework for Shared Control and Multi-Limb Interaction in Virtual Reality

2026-01-01T00:00:00Z

Synergistic Embodiment: A Framework for Shared Control and Multi-Limb Interaction in Virtual Reality Zhou, Hongyu Collaboration is moving beyond shared screens and virtual spaces toward shared bodily action. As VR, robotics, and AI-mediated systems reshape how people work, learn, care, play, and act at a distance, virtual and augmented bodies are becoming shared, programmable media for coordination. They may be co-controlled by multiple users, extended through supernumerary limbs, experienced from distributed viewpoints, or partly governed by semi-autonomous processes. These systems unsettle a foundational assumption in HCI and VR: that embodiment belongs to a single user, body, and perception-action loop. I term this emerging condition collaborative embodiment. This thesis investigates collaborative embodiment as a unified research problem. I first surveyed 137 studies to map methods in collaborative VR and identify gaps in shared control, perspective, and limb augmentation. I then designed and evaluated three VR systems: CoplayingVR for shared hand control, One Body, Two Minds for dynamic perspective switching, and Juggling Extra Limbs for coordination with semi-autonomous virtual arms. Using mixed methods, I show that shared control can improve novice performance and engagement when coordination costs are managed; flexible perspectives balance awareness, comfort, and embodiment; and increasing limb autonomy shifts users from manipulation toward delegation and supervision. This thesis advances Synergistic Embodiment as a conceptual lens and introduces the Interaction Elements Matrix as a design framework for control mapping, perspective strategy, feedback, and autonomy.

Microplastic Mobility and Retention in Geosynthetic Clay Liners

2026-01-01T00:00:00Z

Microplastic Mobility and Retention in Geosynthetic Clay Liners Gao, Yifei Microplastics (MPs) are plastic particles smaller than 5 mm which have recently emerged as environmental contaminants of concern. Landfills serve as major reservoirs of plastic and potential sources of MP pollution due to the accumulation and degradation of plastic waste. MPs have been detected in landfill leachate, with concentrations exceeding 30,000 pieces per litre, well above levels typically observed in natural water bodies. Engineered barrier systems in landfills typically rely on geosynthetic clay liners (GCLs) and an underlying attenuation layer to restrict leachate migration. The swelling behaviour of bentonite in GCLs allows it to achieve low hydraulic conductivity and self-sealing capacity. While GCLs’ ability to prevent the migration of dissolved contaminants has been extensively investigated, their capacity to retain suspended particulate contaminants such as MPs remains poorly understood. This knowledge gap is especially concerning under field-representative conditions, where exposure to wet-dry cycles and calcium-rich leachates is known to reduce bentonite swelling and self-sealing behaviour, potentially compromising long-term barrier performance. An important methodological obstacle in addressing this gap is that conventional laboratory rigid soil columns, typically used to assess contaminant transport, are not suitable for low-permeability soils. This thesis is the first to provide a systematic experimental and numerical investigation of MP transport through GCLs. In all experiments with GCLs hydrated with deionized water, effluent MP concentration remained below the analytical detection limit, and no breakthrough was observed within the test durations, regardless of whether the bentonite was in powder or granulated form. By contrast, in GCLs containing granular bentonite that were exposed to 100 mM CaCl2 and subjected to two wet-dry cycles, early breakthrough was observed for all studied MPs. Includes publication

Explorations in Human Decision Making

2026-01-01T00:00:00Z

Explorations in Human Decision Making Campos, Jerome Sidney Human decision making, even when it involves simple everyday choices has been revealed to have layers of complexity to it in recent years (P. W. Glimcher & Fehr, 2013; Rangel et al., 2008; Rushworth & Walton, 2009). The works contained in this thesis aim at exploring some of those layers, from designing procedures to improve decisions, conducting a meta-analysis on probability weighting, and looking at being hungry and how it affects decisions. In the first chapter, we designed three choice procedures to improve choice quality without reducing the options available or using nudges. Using a within-subject experiment, we tested how our procedures compare to simply picking the preferred option. We find that each of our choice procedures increased the probability that a participant will correctly choose their independently identified highest valued option from the choice set. In the second chapter, we present a meta-analysis of the sensitivity parameter, γ , of probability weighting functions used in models of decision-making under risk. Using 721 parameter estimates from 176 empirical papers, we examined the meta-analytic means, sources of heterogeneity, and domain dependence of γ across gains and losses, as well as when papers did not differentiate between the two (undifferentiated). We found that the meta-analytic mean of γ is significantly below one in all domains and found that estimates are lowest for gains, highest for losses, with undifferentiated estimates lying between the two. In the third chapter, using the decision noise framework proposed by (Shen et al., 2025) we examined whether short term hunger influences simple food related decision-making through changes in early noise and or late noise. Contrary to our first hypotheses, we found that short term hunger does not reduce early noise, but we did find that it does not affect late noise which supported our second hypothesis. Includes publication

Maneuvering the Sampling Process of Generative Models for Feature Alignment and Accelerated Inference

2026-01-01T00:00:00Z

Maneuvering the Sampling Process of Generative Models for Feature Alignment and Accelerated Inference Dinh, Anh Dung Generative Models (GMs) have become increasingly prominent in both applied and fundamental research, with applications spanning nearly every aspect of daily life. Despite their remarkable success in producing high-quality and realistic outputs, GMs still face fundamental challenges during the sampling process. In many cases, even when using the same model, the generated samples can vary dramatically—some closely align with predefined conditions and exhibit excellent quality, while others deviate significantly and appear poor or unrecognizable. This thesis addresses the central question: How can we control the sampling process to avoid poor outputs and consistently generate high-quality features? We approach this problem through an optimization-based framework. First, we formulate the sampling process as an optimization problem. Then, depending on the specific task, we design different objective functions to systematically improve various aspects of the sampling process across different types of generative models. The proposed methods are benchmarked on multiple baselines, covering both diffusion-based and autoregressive generative models. Our results demonstrate that the optimization-based perspective provides a unified and effective approach for enhancing the quality and consistency of samples across diverse generative modeling paradigms.

RDSplat: Robust Watermarking Against Diffusion Editing for 3D Gaussian Splatting

2026-01-01T00:00:00Z

RDSplat: Robust Watermarking Against Diffusion Editing for 3D Gaussian Splatting Zhao, Longjie 3D Gaussian Splatting (3DGS) has established itself as a leading representation for novel view synthesis, underpinning a rapidly expanding ecosystem of commercially valuable digital assets. The growing circulation of these assets has brought IP protection to the forefront, yet the research landscape remains fragmented. This thesis contributes to 3DGS IP protection along two complementary axes. The first contribution is a systematic survey organising 24 existing methods under a unified bottom-up framework with three layers: Gaussian-based perturbation mechanisms; IP protection tasks spanning passive paradigms (watermarking, steganography, tampering localisation) and active paradigms (editing safeguard); and AIGC-era robustness threats. A central finding is that diffusion-based generative editing is a severe and largely unaddressed vulnerability, as it performs iterative semantic reconstruction that suppresses high-frequency content, effectively erasing embedded watermarks. The second contribution targets this vulnerability with RDSplat, a 3DGS watermarking framework robust to both conventional distortions and diffusion-based editing, by anchoring signals in low-frequency spectral components. Carriers are selected via a visibility-aware Mip Score, restricting perturbations to the lowest-frequency 30% of primitives. Training uses a surrogate exploiting the spectral equivalence between Gaussian blur and diffusion editing, avoiding costly generative models in the loop. Recovery is performed by GeoMark, a geometry-aware ViT-S/16 decoder. Evaluation on Blender, LLFF, Mip-NeRF 360, and Instruct-NeRF2NeRF shows RDSplat outperforms the strongest baseline by +0.17 in average bit accuracy under diffusion-based attacks, confirming low-frequency embedding as an effective strategy for 3DGS IP protection in the AIGC era. Includes publication

Exploring Multi-Modal Representation Learning for Embodied Agents

2026-01-01T00:00:00Z

Exploring Multi-Modal Representation Learning for Embodied Agents Yin, Zhenfei Building intelligent embodied agents requires representations that support grounding, prediction, reasoning, and action. Existing systems show a gap between semantic understanding and embodied competence: they describe scenes and generate futures, yet fail to localize objects, reason over geometry, coordinate agents, or execute actions reliably. Representation learning is a fundamental bottleneck for embodied intelligence. This thesis investigates multi-modal representation learning for embodied agents. We study instruction-tuned multi-modal models across images, point clouds, and videos, showing semantic alignment emerges earlier than geometric grounding. We introduce evaluation protocols probing robustness, hallucination, and failure modes, and explore parameter-efficient adaptation strategies preserving grounding across tasks. We examine predictive models as world representations for embodied generalization. Visual realism alone is insufficient; effective world models must integrate multi-modal structure and physical constraints. Through evaluation combining perceptual assessment and embodied execution, models grounded in multi-modal representations—linking vision, geometry, language, and dynamics—generalize more reliably than unguided approaches. Long-horizon reasoning emerges from systems organizing predictive and control skills into adaptive multi-agent structures. Reward-driven self-organizing frameworks enable task decomposition and coordination. Structured constraints and hierarchical supervision show compositional multi-modal representations enable safer, more generalizable embodied behavior. Embodied intelligence is fundamentally a representation and system design problem. Competence emerges when multi-modal representations are task-conditioned, physically grounded, and compositional. These findings provide guidance for building general-purpose embodied agents. Includes publication

Navigating in-betweenness: The everyday politics of queer survival and recognition in Vietnam

2026-01-01T00:00:00Z

Navigating in-betweenness: The everyday politics of queer survival and recognition in Vietnam Tran, Thuy Duong This thesis examines Vietnamese queer lives through the analytical lens of in-betweenness — the condition of navigating incompatible normative orders where satisfying one violates another. Based on fourteen months of ethnographic fieldwork across Vietnam, it argues that in-betweenness is not a transitional state towards Western models of queer visibility and rights, but a permanent one embedded in the local practicalities of queer survival. The thesis analyses four structural positions where in-betweenness operates differently. First, it analyses how LGBTQI+ activists negotiate international donor demands for visible activism against Vietnamese societal demands for restraint, showing how funded recognition enables certain forms of activism but not others. Second, it analyses PFLAG parents who operate between activist infrastructures and kinship obligations, achieving change through care-based strategies that extend rather than by challenging family norms. Third, it analyses working-class queer individuals who practise 'dependent inclusion', or the over-performance of care and economic obligations to secure family belonging, revealing the double burden of building independent lives while maintaining affective family ties. Fourth, it analyses how poor trans women excluded from activist circles and family welfare produce lô tô, a form of collective survival through performative art that addresses pain without eliminating it. These four positions foreground the hard navigational work of a visibility politics where the liberal demands of international funders meet the relational demands of local queer lives, producing remarkable achievements in legal reform, media representation and community mobilisation but revealing the inherent limitations of funded recognition for people who have neither escaped the patriarchal norms of their society nor embraced the liberal norms of their foreign supporters. Includes publication

Negotiating Postcolonialism through Whiteface: Performances of Whiteness in Contemporary Senegal and Australia

2026-01-01T00:00:00Z

Negotiating Postcolonialism through Whiteface: Performances of Whiteness in Contemporary Senegal and Australia Okkes-Sane, Charlotte This thesis examines the use of “whiteface” – the mimicry of Whiteness in theatre, television and writing – by Senegalese and Aboriginal and Torres Strait Islander artists as a tool for anticolonial critique. I argue that whiteface often disturbs and challenges colonial Whiteness, but also simultaneously reifies it. Across contexts as diverging as Senegal and Australia, artists resist postcolonial norms upheld by colonialism: the violence of migration bureaucracy, linguistic hierarchies which depreciate local languages, literary repetitions of colonial representations of Indigenous peoples, and the dehumanisation of the colonial gaze. Whiteface becomes an effective tool (for both the artist and the reader/audience) to understand and critique colonial violence. None of these artists, however, are able to fully extricate themselves from complicity with colonial Whiteness. To a certain extent, their work remains tethered to the colonial world through, for example, collaboration with organisations with their own agendas in place, self-censorship, and the reinscription of colonial literary tropes. Thus, whiteface texts demonstrate the complexity of postcolonial critique for marginalised communities, who voice their ideas whilst at the same time trying to survive (economically, artistically) in colonial societies.

From Jets to Mergers: Multiwavelength Signatures of Synchrotron Transients

2026-01-01T00:00:00Z

From Jets to Mergers: Multiwavelength Signatures of Synchrotron Transients Gulati, Ashna Synchrotron transients arise from explosive astrophysical events that launch relativistic outflows into their surrounding environments. Their interaction then produces synchrotron emission which further probes the physics, geometry, and environments of some of the most energetic phenomena in the Universe. This thesis investigates relativistic outflows using wide-field radio surveys and coordinated multi-wavelength follow-up observations. Late-time radio observations with the Australian SKA Pathfinder (ASKAP) are used to search for afterglows associated with eleven well-localised gravitational-wave events detected by LIGO/Virgo. No convincing radio counterparts are found out to approximately 1500 days after merger. These non-detections constrain the inclination angles, circummerger densities, and energetics of relativistic jets, and define the conditions under which late-time radio emission from compact binary mergers can be detected in wide-field surveys. This thesis also presents the discovery and characterisation of ASKAP J0055−2558, a long-lived luminous extragalactic radio transient identified during a search for orphan afterglows. Broadband observations reveal an evolving synchrotron spectrum and a mildly relativistic outflow consistent with either an off-axis long gamma-ray burst afterglow or an off-nuclear tidal disruption event, demonstrating the ability of radio surveys to uncover rare classes of transients. Finally, a radio-to-X-ray study of the afterglow of GRB 240825A reveals evidence for a two-component jet and a hard electron energy distribution ($p<2$), providing new constraints on the structure and microphysics of relativistic jets. Together, these studies demonstrate the power of combining wide-field radio surveys with broadband follow-up observations to constrain the physics, geometry, and environments of explosive astrophysical outflows, establishing radio surveys as a critical tool for exploring the time-domain Universe. Includes publication

Hydrogen Trapping and Embrittlement Behaviours in Steels

2026-01-01T00:00:00Z

Hydrogen Trapping and Embrittlement Behaviours in Steels Huang, Chao Hydrogen embrittlement is a major challenge for the durability of structural steels used in the emerging hydrogen economy. This thesis investigates how hydrogen interacts with microstructural defects, how these interactions contribute to embrittlement, and how they can be mitigated through microstructural design, with the aim of establishing microstructure–property relationships in hydrogen-containing environments (Chapters 1 and 2). Using advanced characterization techniques and principles of metal physics, this work reveals how defects such as dislocations, interfaces, and precipitates act as hydrogen traps and influence embrittlement behavior. The concepts of weak and strong hydrogen traps are explored, and a new experimental strategy is introduced to distinguish their respective roles (Chapters 4–6). By correlating atom probe tomography, transmission-based microscopy, and local mechanical characterization, hydrogen distributions at specific defects are directly linked to deformation behavior, providing an atomic-scale understanding of hydrogen–defect interactions (Chapter 7). The results further demonstrate that strong hydrogen traps can be intentionally activated through microstructural optimization, improving resistance to embrittlement (Chapter 8). In addition, this work examines interstitial hydrogen solution behavior in face-centered cubic alloys, with the results showing good agreement with theoretical predictions regarding hydrogen affinity. These findings provide new insights into hydrogen partitioning and alloy design. Finally, future directions are proposed for atomic-scale hydrogen mapping using cryogenic APT and for the development of next-generation hydrogen-resistant alloys for advanced energy applications (Chapters 9 and 10).

ADVANCED ALGORITHMS FOR RIS-ASSISTED WIRELESS COMMUNICATION SYSTEMS

2026-01-01T00:00:00Z

ADVANCED ALGORITHMS FOR RIS-ASSISTED WIRELESS COMMUNICATION SYSTEMS Li, Weijie Reconfigurable intelligent surface (RIS) has emerged as a promising technology for 6G wireless networks, offering the ability to reshape propagation environments and improve communication performance. Despite its potential, deploying RIS faces challenges in channel estimation, signal processing, and system integration. This thesis develops novel algorithmic solutions for RIS-assisted wireless communication systems, focusing on improving estimation accuracy, reducing complexity, and enabling integrated sensing and communication. First, we develop UAMP-SBL-PCI, exploiting structured sparsity in RIS-assisted channels. By combining partial common support identification with unitary approximate message passing, it reduces complexity while improving accuracy. Simulations demonstrate excellent performance across various environments. Second, we propose a unified framework for ULA RIS systems jointly performing user positioning and channel estimation. A vision transformer (ViT) enables adaptive positioning, while a position-based dictionary design reduces dictionary size and resolves the off-grid problem. A modified sparse Bayesian learning algorithm with early stopping prevents overfitting for static deployment scenarios. Third, we extend this to dynamic environments with a three-stage joint channel estimation and positioning framework for UPA RIS systems, employing a graph attention network (GAT) for robust positioning, location-aware dictionary refinement, and meta-learning for rapid adaptation. An iterative refinement mechanism further reduces pilot overhead. These contributions span from basic channel estimation to advanced integrated sensing and communication, enabling practical RIS implementation with reduced costs and improved accuracy for 6G systems.

Robust Implied Volatility Surface Construction via B-Spline-Enhanced Graph Neural Operators

2026-01-01T00:00:00Z

Robust Implied Volatility Surface Construction via B-Spline-Enhanced Graph Neural Operators Ye, Xuan The reliable recovery of implied volatility surfaces from sparse option quotes remains challenging, yet it is essential for pricing, hedging, and market risk measurement. In practice, widely-used parametric specifications can be fast and arbitrage-aware, but their stability depends on repeated calibration and can deteriorate when market data are irregular. Operator learning offers a complementary perspective by modelling a mapping from scattered observations to a continuous surface, though many implementations become expensive because they predict values over dense grids and therefore require large training sets. This thesis develops a graph-based neural operator that targets a low-dimensional spline representation of the surface. A bivariate cubic B-spline is first used to obtain an initial smooth approximation, after which the operator learns residual adjustments to the spline coefficients rather than producing a full grid of volatilities. This coefficient-level formulation embeds smoothness by construction and reduces computational burden. In addition, the analytic derivatives available under the spline representation enable butterfly and calendar no-arbitrage penalties to be imposed without finite-difference noise. Experiments on daily S&P 500 index options show that the resulting framework is more stable under sparse conditions and achieves improved efficiency relative to common parametric baselines and grid-based neural operator models. These results support spline-prior operator learning as a practical route to constructing smooth, arbitrage-regularised volatility surfaces.

Graph-Based Machine Learning for Real-World Applications

2026-01-01T00:00:00Z

Graph-Based Machine Learning for Real-World Applications Yan, Kuan Graph-based machine learning has emerged as a powerful approach for modeling complex real-world data, particularly in finance and medical research, where data are heterogeneous, noisy, incomplete, and highly structured. This thesis investigates graph-based machine learning methods for addressing practical challenges in financial and medical AI, with an emphasis on robustness, interpretability, and fairness. The first part of the thesis focuses on financial applications. A heterogeneous graph learning framework is developed for credit card fraud detection by modeling complex transactional relationships. Feature importance–based edge weighting is incorporated into a graph neural network to improve predictive performance and model interpretability. The thesis further investigates fairness in graph neural networks by proposing a framework that balances group fairness and individual fairness, demonstrating that these objectives can be jointly improved in graph learning tasks. The second part of the thesis focuses on medical AI for retinal disease research. Machine learning methods are applied to predict subretinal lesion severity and identify genes associated with disease progression from high-dimensional transcriptomic data. Building on this work, a graph pseudotime analysis framework is proposed to infer disease progression trajectories, while neural stochastic differential equations are employed to characterize pathway dynamics and identify transition points. Overall, this thesis demonstrates that graph-based machine learning provides a flexible and effective framework for modeling complex data across financial and biomedical applications. The proposed methods improve predictive performance, interpretability, and fairness while providing biologically meaningful insights into disease progression. Includes publication

Lattice-Boltzmann Large Eddy Simulation Solver for Atmospheric Transport and Dispersion Applications

2026-01-01T00:00:00Z

Lattice-Boltzmann Large Eddy Simulation Solver for Atmospheric Transport and Dispersion Applications Waters, Brendan Ashley Accurate prediction of metropolitan climate conditions is essential for public health and national security, directly influencing commercial, environmental, social, and political activity. High-fidelity dispersion modelling therefore has broad utility, including risk assessment, urban planning, emergency response, and sustainable city management. However, urban morphologies are highly heterogeneous, producing inherently unsteady wind fields with turbulent processes spanning wide spatial and temporal scales. Consequently, industrial-scale simulations involve many degrees of freedom, making prognostic modelling demanding. This thesis presents a wall-modelled large eddy simulation solver for atmospheric transport and dispersion in urban environments at grid resolutions of 1–10m. A hybrid method is considered, wherein the hydrodynamics are modelled using a D3Q27 cumulant lattice Boltzmann method (LBM), while passive scalar transport is modelled via the MUSCL-Hancock finite volume scheme. Following a brief literature review and introduction of the numerical methods, this work systematically the spectral bandwidth of various cumulant LBM-LES formulations, and wall-modelling techniques. The fully integrated solver is then applied to validation cases in both idealised urban environments using the Mock Urban Setting Test database and real-world geometries using the Joint URBAN 2003 database for downtown Oklahoma City. Results show that the WMLES solver predictions are in good agreement with experimental data, exceeding industry-standard performance criteria. In the idealised urban canopy, velocity predictions achieve FAC1.3 accuracy. Concentration statistics exhibit larger deviations but remain within experimental uncertainty. For downtown Oklahoma City, the solver retains strong predictive capability in real urban conditions; although formal accuracy is slightly below FAC1.3 despite minor geometric discrepancies, it remains above the industry-standard FAC2 threshold. Includes publication

Deadline Scheduling with Predictions

2026-01-01T00:00:00Z

Deadline Scheduling with Predictions Zhang, Luoyuan Online deadline scheduling has been extensively studied. Most prior works assume a known deadline per job, but this is sometimes impractical, e.g., submission and withdrawal of cloud jobs. Deadline scheduling with unknown deadlines for real-time systems is a challenging problem, where deadlines need to be predicted. This thesis considers the problem of maximizing the number of early jobs under unknown dead lines. In this setting, a job has an unknown deadline, which becomes known only when it is reached, and the job immediately fails after that. The study of this prob lem in this work is within the learning-augmented framework, where the scheduler can access deadline predictions, which may come with errors. A simple error metric is defined to quantify prediction quality and design an algorithm with a competitive ratio as a function of the prediction error. The algorithm achieves performance matching the optimal online scheduler that knows the deadlines and smooth performance degradation as the prediction error increases. Simulations are run to compare the performance of the proposed algorithm with the existing ones. Simulations show that the performance of the proposed algorithm consistently exceeds the conservative theoretical bound; thus, the bound represents a pessimistic view of the algorithm; one can expect exceptional practical performance.

Sounding the Wallacea: The Siboga Oceanographic Expedition (1899-1900)

2026-01-01T00:00:00Z

Sounding the Wallacea: The Siboga Oceanographic Expedition (1899-1900) Maharani, Annissa Ayu This thesis examines the historical context of the research on ocean circulation in the eastern Indonesian waters conducted by the Dutch Siboga Expedition (1899-1900). The region is characterised by abundant and scattered islands, while the seafloor exhibits countless elevations and depressions in close proximity. In the late 19th century, this complex underwater topography in the eastern Indonesia was believed to yield a distinctive physical environment to which marine fauna would respond. The Siboga aimed to investigate deep-layer water communication across basins and to speculate about the effects on the distribution of marine fauna. The thesis emphasises the role of nautical charts, as both scientific instruments and navigational guidance, which were often inaccessible to scientists due to their association with the colonial administration and top-level authorities. It highlights the involvement of naval powers and the mobilisation of the colonial maritime infrastructure in sustaining oceanographic expeditions within the archipelagic colony. The thesis argues that the Siboga made significant contributions to the understanding of physical oceanography in Indonesian waters, producing foundational observations on stratified deep circulation between the neighbouring Pacific and Indian Oceans. It illustrates how knowledge is produced under multi-layered states of technological, logistical, and political factors, including manual sounding technology, changing monsoonal atmospheric conditions, and colonial hydrographic records.

Evaluating The Impact Of The Speech-Language Pathology Primary Contact Model Of Care In Diagnostic Assessment Of Voice Disorders.

2026-01-01T00:00:00Z

Evaluating The Impact Of The Speech-Language Pathology Primary Contact Model Of Care In Diagnostic Assessment Of Voice Disorders. Payten, Christopher Letton Voice disorders affect millions and are costly. Early assessment is vital, but access to ENT and multidisciplinary clinics is limited. The speech-language pathology primary contact (SLP-PC) model offers quicker evaluation before or instead of ENT assessments, though its reliability and diagnostic contribution need more evidence. This thesis presents five studies on SLP-PC in adult VD diagnosis: an observational cohort, a literature review, a global MDT survey, a GP referral review, and another cohort study, examining outcomes, frameworks, practice patterns, predictors, and diagnostic agreement. Integrating SLP-PC into an ENT MDT pathway reduced wait times by an average of 277 days compared to traditional pathways. Most patients (81%) managed by SLP didn't need ENT assessment, but 7% of urgent cases required ENT. SLP-PC was estimated to cut staffing costs by 27%. A review of 20 frameworks across 2,675 publications highlighted the need for clearer classification systems for VDs suitable for first-line SLP intervention, a need that was subsequently addressed. A survey of 109 clinicians showed growing support for diverse diagnostic pathways, including SLP-PC, emphasising the importance of case history. An analysis of GP referrals found no strong predictors of assessment approach, underscoring the need for better triage information and GP training. The SLP-PC telehealth model showed promise in predicting urgent ENT needs, with substantial diagnostic agreement. Overall, these studies demonstrate that SLP-PC is a reliable, effective model that improves access and diagnostic efficiency while enhancing clinical skills among ENT and SLPs within an integrated MDT VD service. This research advances global voice assessment practices, supports SLP-first pathways, offers insight for GPs, and highlights the need for standardised classification systems. Further research is needed to understand the diagnostic reasoning of SLPs and ENTs in VDs and the economic impact of SLP-PC. Includes publication

Investigating the therapeutic potential of Australian honey as a prebiotic and anti-inflammatory to improve gut health

2026-01-01T00:00:00Z

Investigating the therapeutic potential of Australian honey as a prebiotic and anti-inflammatory to improve gut health Schell, Kathleen Rose The gut microbiome is a complex microbial community that plays a critical role in host health. Diet is a major determinant of its composition and function, while dysbiosis is associated with intestinal inflammation and numerous chronic diseases. Australian honey has emerged as a potential functional food due to its complex carbohydrate content and reported antimicrobial, prebiotic and anti-inflammatory properties. This thesis investigated the antibacterial, antioxidant, prebiotic and anti-inflammatory potential of 56 Australian honeys from diverse floral sources using in vitro, ex vivo and in vivo models. Australian honeys exhibited minimal antibacterial activity under anaerobic conditions, indicating little potential to adversely affect beneficial gut bacteria, while antioxidant capacity differed significantly between floral types. Ex vivo studies demonstrated that all honeys promoted Lactobacillaceae growth and enhanced suppression of opportunistic pathogens in mixed faecal cultures. In vivo, Australian honey increased the abundance of beneficial taxa, including members of the Lactobacillaceae and Muribaculaceae families, in mice fed a standard chow diet. Following western diet-induced gut microbiome restructuring, honey promoted alternative carbohydrate-metabolising taxa rather than restoring those reduced by the diet. In a dextran sodium sulphate (DSS) model of colitis, honey reduced histopathological indicators of inflammation and promoted immune phenotypes associated with inflammation resolution. Collectively, these findings demonstrate that Australian honey possesses antimicrobial, antioxidant, prebiotic and anti-inflammatory properties that support its potential to beneficially modulate the gut microbiome and attenuate intestinal inflammation. These results provide a foundation for future investigations into Australian honey as a functional food for gastrointestinal health. Includes publication

DEVELOPING PROTEIN-RICH FOODS FROM AUSTRALIAN GROWN FABA BEANS

2026-01-01T00:00:00Z

DEVELOPING PROTEIN-RICH FOODS FROM AUSTRALIAN GROWN FABA BEANS Hopf, Andreas Nikolaus This thesis examined dry fractionated faba bean protein concentrate through systematic investigation of extraction method effects, controlled modification strategies, and application in diverse food structuring systems. Initial characterisation compared dry fractionated and wet fractionated pulse proteins from faba bean, mung bean, chickpea and red lentil against commercial soy protein concentrate. Extraction process critically impacted functional properties and composition. Controlled thermal treatment applied prior to dry fractionation altered functional behaviour of dry fractionated faba bean concentrate. Partial denaturation of the proteins modified gelation mechanism. Thermal treatment mitigated the reduced water-holding capacity associated with elevated Calcium Sulphate coagulant concentrations, revealing synergistic effect of thermal pretreatment and gelation. The gelation properties of faba bean protein were further investigated in the application of tofu-like bean curds. Coagulant selection in faba bean tofu production was critical. Finally, the assessment of dry fractionated faba bean protein under high moisture extrusion conditions using a micro compounder explored the material for fibrous meat analogue products. This research demonstrates that dry fractionated faba bean protein concentrate possesses distinctive functional characteristics. These characteristics can be systematically enhanced through pretreatments, including thermal treatment. The protein concentrate supports implementation in both conventional gel-based and emerging extrusion-based food applications. These outcomes advance industrial adoption pathways for sustainably produced pulse proteins while delivering practical formulation frameworks for manufacturers pursuing alternatives to animal-based and wet fractionated protein ingredients. The findings hold particular relevance to the Australian pulse industry, by exploring pathways to valorise local protein-rich crops for the food industry. Includes publication

Inflammation and its Effects on Plaque Cell Fate in Atherosclerosis

2026-01-01T00:00:00Z

Inflammation and its Effects on Plaque Cell Fate in Atherosclerosis Lin, Alexander Atherosclerosis, a chronic inflammatory disease of the arteries, causes myocardial infarction and stroke when an unstable atherosclerotic plaque ruptures. Stable plaques consist of a thick fibrous cap, composed of abundant collagen-secreting cells, with fewer inflammatory infiltrates, whereas unstable plaques have a thin fibrous cap with few collagen-producing cells and excessive inflammatory cells. Vascular smooth muscle cells (SMCs) are traditionally considered the primary source of collagen production within the lesion, but studies over the last decade have revealed their capacity to also modulate their phenotype and worsen lesion stability instead. However, the cellular mechanisms which regulate SMC phenotype are not fully understood. This thesis aims to investigate how modulating inflammation, using colchicine as an anti-inflammatory therapy and diabetes as a pro-inflammatory cardiovascular risk factor, influences SMC phenotypic behaviour within the atherosclerotic lesion, and how this ultimately affects plaque stability. First, we demonstrate that colchicine directly promotes a contractile SMC phenotype, an effect which appears independent of its canonical anti-inflammatory properties in immune cell populations. Second, we show that diabetes induces the expansion of multiple SMC clones into the lesion but hinders their trans-differentiation to a fibroblast-like phenotype, ultimately leading to declines in lesion stability. Third, we find that colchicine intervention in diabetic lesions did not alter the stability of plaques and had very little influence on SMC behaviour. Taken together, these findings highlight the importance of SMCs and their phenotypic transitions in atherosclerosis, both in the therapeutic stabilisation of lesions, and in driving disease pathogenesis with co-morbidities. Includes publication

Knowledge-Driven Visual Reasoning Under Data Scarcity

2026-01-01T00:00:00Z

Knowledge-Driven Visual Reasoning Under Data Scarcity Zhang, Xu Visual reasoning connects seeing with understanding by requiring vision systems to produce structured, decision-relevant outputs from pixels. These capabilities are crucial for open-world deployment and are viewed as a prerequisite for general intelligence. Yet, despite progress driven by deep learning and pretraining, state-of-the-art visual reasoning pipelines remain constrained by a fundamental bottleneck: they rely heavily on structured supervision. Reasoning-oriented tasks require annotations such as boxes, which are more expensive to obtain than class labels, making data-hungry fine-tuning difficult to extend to specialized applications. This thesis advances data-efficient visual reasoning from the perspective of knowledge-driven learning. Its unifying principle is that when structured labels are scarce, missing supervision can be compensated by making relevant knowledge explicit and using it to guide learning and inference. Here, knowledge is instantiated as explicit guidance signals, including linguistic knowledge, reasoning priors, and exemplar-derived textual conditions. We develop it across increasingly stringent scarcity regimes. First, under intra-domain sparsity with large appearance variability, we improve pose estimation via contrastive alignment that leverages linguistic knowledge to enhance keypoint correspondence learning. Second, for cross-modal reasoning, we improve label efficiency in referring object detection by integrating interpretable priors into transformer-based pipelines. Third, under the regime where data scarcity is compounded by severe domain shift, we propose a training-free paradigm that uses language as a domain-invariant bridge, converting exemplars into textual conditions that guide frozen detectors. Extensive experiments show that explicit knowledge guidance can serve as a practical surrogate for structured supervision, offering a scalable route for applying foundation models to data-scarce, real-world visual reasoning. Includes publication

Automated Radiation Dose Monitoring in Computed Tomography: A Comprehensive Evaluation of Clinical Effectiveness Adoption and User Experience

2026-01-01T00:00:00Z

Automated Radiation Dose Monitoring in Computed Tomography: A Comprehensive Evaluation of Clinical Effectiveness Adoption and User Experience Alanazi, Mohammed Aim: This thesis investigated the role of automated dose monitoring systems (DMS) in computed tomography (CT) radiation dose monitoring and optimisation. Method: A multi-method research design comprising two systematic reviews and three empirical studies was undertaken. The first review evaluated the use of DMS high-dose alert functions and their impact on CT dose optimisation, while the second examined DMS applications, benefits, and challenges in CT practice. The third study surveyed CT dose monitoring practices and DMS adoption in Australian radiology facilities. The fourth study explored the experiences of medical physicists and radiographers using DMS in CT. The final study retrospectively assessed a DMS's ability to detect high-dose CT events at an Australian radiology facility. Results: Study 1 showed that DMS systems are valuable for identifying high-dose events and supporting dose optimisation. Study 2 showed that DMS systems facilitate benchmarking, tracking and estimation, although challenges related to data inconsistencies and integration issues were reported. Study 3 found that while most Australian radiology facilities performed CT dose assessments, these were commonly undertaken annually using traditional methods, and 41% used DMS. Study 4 showed that DMS users reported improved workflow efficiency, CT dose assessment, and radiation protection practices, although challenges related to system setup, data validation and IT infrastructure were identified. The final study found that 851 of 11,865 CT examinations (7%) triggered DMS alerts, most commonly in spine, joint, and abdomen–pelvis CT exams, with contributing factors related to patients and operators. Conclusion: Automated DMS systems support CT dose monitoring and optimisation through continuous dose assessment, benchmarking, and identification of high-dose events. Wider adoption, supported by appropriate implementation and training, would strengthen radiation protection and enhance patient safety. Includes publication

Multi-Scale Thermo-Mechanical and Microstructural Simulations of Wire Arc Additive Manufacturing (WAAM) of 316L Stainless Steel

2026-01-01T00:00:00Z

Multi-Scale Thermo-Mechanical and Microstructural Simulations of Wire Arc Additive Manufacturing (WAAM) of 316L Stainless Steel Valiente Dies, Fernando Wire Arc Additive Manufacturing (WAAM) of 316L stainless steel offers a practical route to large structural components, yet industrial uptake is limited by distortion, residual stresses, and microstructure heterogeneity, together with a lack of experimentally validated simulation workflows. This thesis develops and validates an integrated framework that links WAAM process parameters to thermo-mechanical response, hardness, and microstructure. Purpose-built benchmark walls and T-joints were designed and fabricated to generate co-registered datasets: in-process thermocouples, full-field distortion by GOM, microhardness maps, EBSD textures, and internal residual stresses from neutron diffraction and the contour method. A coupled thermo-mechanical model in MOOSE, with temperature-dependent properties, and progressive element activation, was calibrated to the thermal data and validated against distortion and hardness. A hardness surrogate tied to accumulated equivalent plastic strain, reproduced measured hardness gradients between wall, HAZ, and base plate. For T-joints, simulations captured the sign, magnitude, and spatial distribution of tensile and compressive regions observed by diffraction and contour measurements to clarified their evolution through deposition and cool-down. Microstructure was interrogated with phase-field and Cellular Automata models to reproduced epitaxial columnar growth, grain size trends through height, and texture consistent with EBSD. The outcome is an open, experimentally anchored pathway for pre-build what-if analysis and process design in WAAM 316L. The work contributes: a shared benchmark dataset, a calibrated modelling stack connecting temperature, stress, distortion, and hardness, microstructure-informed interpretation of anisotropy, and actionable guidance on energy input, interpass temperature, bead sequencing, and restraint. The proposed framework is readily extensible to more complex geometries and other alloys. Includes publication

On Vassiliev Invariants and Weight Systems of Classical and Welded Knots

2026-01-01T00:00:00Z

On Vassiliev Invariants and Weight Systems of Classical and Welded Knots Lin, Damian Vassiliev invariants are a special class of knot invariant that are analogous to polynomial functions on, for example, the real line. We review how Vassiliev invariants can be constructed from Lie algebra objects in arbitrary monoidal categories and are completely determined by their weight systems. We make some computations of the values of the g2 and f4 exceptional Lie algebra weight systems on a certain special class of chord diagrams. We generalise a construction of Hinich-Vaintrob to welded knots, thereby constructing a universal welded weight system.

Efficient Edge-AI: Towards the Future of Implantable and Smart Medical Devices

2026-01-01T00:00:00Z

Efficient Edge-AI: Towards the Future of Implantable and Smart Medical Devices Herbozo Contreras, Luis Fernando Epilepsy affects over 1% of the global population and imposes substantial clinical, social, and economic burdens. Although pharmacological therapy is the first line of treatment, approximately 30–40% of patients remain drug-resistant, making neuromodulation one of the few viable options. However, current neuromodulation systems are largely open-loop or depend on cloud-based AI, limited by latency, power, privacy, and scalability. These constraints hinder autonomous, personalised, implantable closed-loop neurostimulation. This thesis investigates neuromorphic computing as a paradigm for next-generation closed-loop neuromodulation, focusing on seizure detection and prediction in epilepsy. It introduces neuromorphic neuromodulation and shows how biologically inspired, on-device intelligence can support self-responsive and personalised therapies. Building on this framework, the thesis develops seizure detection systems using liquid-time constant neurons and dendritic spiking mechanisms with heterogeneous temporal dynamics. These models enable efficient neural-signal representation without expensive feature extraction and show robust out-of-sample generalisation across patients and recording conditions on large-scale clinical EEG datasets. The thesis also addresses edge learning by proposing a neuromorphic learning rule for on-device adaptation toward patient-specific treatment under strict power and memory constraints. This mechanism can be adopted by the developed models to support personalised, low-latency intelligence at the edge. Finally, learnable activation functions are explored within artificial neural networks to improve training efficiency and interpretability, highlighting a pathway for integration with neuromorphic technology. Together, these contributions advance neuromorphic neurotechnology toward autonomous, personalised, and continuously learning closed-loop systems, with implications beyond epilepsy to a broader range of neurological disorders.

Blockchain-based Advanced Information Infrastructure and its Applications in Demand-Side Energy Systems

2026-01-01T00:00:00Z

Blockchain-based Advanced Information Infrastructure and its Applications in Demand-Side Energy Systems Yu, Teng The energy sector is moving from centralised one-way power systems to decentralised bidirectional smart grids driven by distributed energy resources (DERs). This shift requires infrastructure for secure coordination, verifiable markets, and prosumer data privacy. Existing blockchain systems have limited scalability, security, and interoperability, restricting high-frequency smart-grid applications. This thesis designs blockchain infrastructure for peer-to-peer (P2P) energy trading and power load forecasting (PLF). It has three theoretical innovation points. First, a dual-blockchain architecture with an Improved Optimistic Rollup (IOR) improves vertical throughput by offloading heavy computation from a primary to a secondary blockchain. Second, a Transaction Batch Generation (TBG) protocol for leaderless Byzantine Fault Tolerance (LBFT) consensus improves horizontal throughput by letting every node broadcast blocks, avoiding redundant transaction rebroadcast, and reducing transaction censorship to nearly zero. Third, a Blockchain-of-Blockchains (BoB) architecture, with Cross-Chain Token Exchange (CCTE) and Cross-Chain Data Interoperability (CCDI), supports asset and data transfer across blockchains with lower latency and memory overhead. The infrastructure is applied to two demand-side problems. For P2P market clearing, it is combined with Trusted Execution Environments (TEEs), D-TASK, and TEAR-DO to address the privacy-robustness-latency trilemma: prosumer data remain private, distributed optimisation preserves optimal convergence, and clearing time is reduced by an order of magnitude over state-of-the-art methods. For PLF, CTP-FL combines the infrastructure with message coding and commitment-based dual consensus to preserve local-model privacy, tolerate Byzantine faults at client and server sides, and keep communication latency below local model training time. Includes publication

Towards Geometry-Grounded World Understanding

2026-01-01T00:00:00Z

Towards Geometry-Grounded World Understanding Tang, Liyao Understanding the three-dimensional (3D) world underpins embodied artificial intelligence, enabling robotics, autonomous driving, augmented and virtual reality, and, more broadly, spatial intelligence. Yet the most direct representation of a 3D scene can be deceptively simple: a point cloud, \ie, a set of Cartesian coordinates sampled from scene surfaces, from which objects and semantic structure must be inferred. In practice, point clouds are noisy, incomplete, and irregularly sampled, making reliable interpretation fundamentally challenging. Geometry-grounded world understanding demands semantics that are consistent with metric 3D geometry. Semantic segmentation provides a principled route from geometric measurements to high-level scene understanding. In 3D point clouds, semantic segmentation anchors geometry-grounded world understanding by coupling fine-grained semantics with explicit 3D geometry and confronting a core perceptual challenge: forming structured and coherent interpretation from noisy, incomplete, and irregular samples. This thesis investigates how explicit 3D geometry can be elevated from a passive input to a source of structure for learning and generalization. We treat geometry as a prior that constrains what a plausible segmentation should look like, modulates how noisy supervision should be used, and guides how models adapt when spatial statistics shift. Building on this view, we advance geometry-grounded scene segmentation along three complementary aspects of the learning problem: the output, the supervision, and the adaptation Overall, this thesis approaches scene segmentation by structuring the outcomes of learning, the supervision that guides learning, and the contexts in which learning occurs, all grounded in explicit 3D geometry. In doing so, we advance a geometry-grounded view of world understanding in which explicit 3D geometry shapes both learning and generalization.

Effects of the Post-Heat Treatment (Furnace Cooling) on the mechanical strength and dimensional accuracy of 3D printed PEEK in FDM method

2026-01-01T00:00:00Z

Effects of the Post-Heat Treatment (Furnace Cooling) on the mechanical strength and dimensional accuracy of 3D printed PEEK in FDM method Deng, Yunxiang In recent decades, the production of polymeric parts using fused deposition modelling (FDM) has gained significant attention in the field, owing to its design flexibility, low cost, and time-efficient prototyping capabilities. Nevertheless, the inherently as-built limitation constrains the performance and challenges the broader applications. To address these limitations, the post-heat treatment or annealing has long been applied as one the critical post processing techniques for enhancing the materials properties and its performance. Despite the beneficial effects of the post-heat treatment on the mechanical strength, its effect on the long-term tribological performance with the involvement of complex structures are limited. While it is often assumed that improvements in mechanical properties lead to enhanced tribological performance, tribological properties are not intrinsic material properties. Instead, they are instead dependent strongly on the specific system and operating conditions in which a material or structure has to function. Among the tribological studies, the friction-induced vibration (FIV) is a critical issue, causing unwanted noise, wear, and potential system failure. Although the proposed active or passive controls can mitigate FIV, they inevitably increase the complexity in the design and implementation of the whole system. The re-entrant auxetic structure was employed in this study as the solution, characterized by Negative Poisson’s ratio (NPR). Notably, the performance of AM-fabricated parts remains highly sensitive to the external environmental stimuli, particularly temperature.

Probabilistic Learning and Inference for Multi-Modal Trajectory Generation and Adaptation

2026-01-01T00:00:00Z

Probabilistic Learning and Inference for Multi-Modal Trajectory Generation and Adaptation Yin, Zeya Robotic systems are fundamental to enhancing efficiency, advancing autonomous processes, and performing tasks that are hazardous or routine in daily life. However, the complexity and unpredictability of various tasks require that robots can adapt effectively to changing environments. This thesis aims to advance the state of the art in robotic motion generation by presenting an ensemble of approaches that combine motion planning, learning from demonstration techniques, and modern probabilistic methods. Our goal is to develop principled frameworks that can produce diverse and reliable multimodal solutions for robotic tasks. Includes publication

Attention Calibration for Reducing Hallucination in Large Vision-Language Models

2026-01-01T00:00:00Z

Attention Calibration for Reducing Hallucination in Large Vision-Language Models Zhu, Younan Large Vision-Language Models (LVLMs) exhibit impressive multimodal reasoning capab- ilities but remain highly susceptible to object hallucination, where models generate responses that are not factually aligned with the visual content. Recent works attribute this issue to an inherent bias of LVLMs where vision token attention map has spurious focus on certain positions, and propose to mitigate this issue by reordering visual tokens. However, we find that different LVLMs exhibit different correlations between attention and spatial position, which makes the existing static solution difficult to generalize to other LVLMs. To begin with, we investigate the attention bias introduced by image tokens through a toy experiment, in which a blank image is fed into the model to capture its position-dependent bias. We then remove this bias from the original attention map, which already leads to a substantial reduction in hallucinations. This proof of concept validates the core intuition behind attention calibration. Building on this insight, we propose Dynamic Attention Calibration (DAC), a lightweight, plug-and-play module that leverages contrastive learning to dynamically enforce positional invariance. Unlike static baselines, DAC adapts to different models and inputs in a robust and learnable manner, offering a generalizable solution to mitigate attention-related hallucinations in LVLMs. Comprehensive experiments across multiple benchmarks demonstrate that DAC significantly reduces object hallucination while improving general multimodal alignment. Our method achieves state-of-the-art performance across diverse LVLM architectures on various metrics.

Design and optimisation of OpenStride an open-source inexpensive force plate actometer

2026-01-01T00:00:00Z

Design and optimisation of OpenStride an open-source inexpensive force plate actometer Yang, Yang Quantitative assessment of rodent motor behaviour is central to preclinical neuroscience, and evaluation of naturalistic movement is of particular interest. In the early 2000s, force plate actometry (FPA) was developed to track rodent subject centre-of-mass (COM) with high temporal and spatial precision, enabling varied quantifications relevant to motor performance and behaviour. While FPA was commercialised, these systems cost ~$15,000 AUD and are no longer produced. Thus, despite clear utility, particularly in movement disorders and neuropsychiatric research, FPA adoption has remained limited. In this work, we sought to develop an open-source, low-cost FPA system called OpenStride, and to characterise how its performance is shaped by hardware and signal processing choices, providing users with information needed to use it effectively across varied experimental contexts. We designed OpenStride to be able to be fabricated for ~$800 AUD using standard 3D printing and laser cutting. OpenStride achieves minimal static drift during long recordings and minimal signal displacement in response to environmental perturbation; further, it reliably tracks position and distance and can separate groups of rodents with established motor phenotypes. We characterised the influence of five key variables on measurement quality: applied load, load cell excitation voltage, mechanical damping, platform mass, and signal filtering, providing suggestions on how to optimise setup in a context-dependent manner. Our findings support the interpretation that OpenStride is capable of meaningful motor and behavioural quantification across rodent species and experimental paradigms. Its performance is influenced by modifiable parameters, and understanding these relationships allows users to configure the system to suit their specific needs. All hardware and software files have been distributed freely via GitHub, with the intent of making force plate actometry accessible to the research community. Includes publication

Multimodal Learning for Multi-scope Brain Lesion Segmentation

2026-01-01T00:00:00Z

Multimodal Learning for Multi-scope Brain Lesion Segmentation Chen, Jiahao Brain lesion-related diseases, including glioblastoma, stroke, and demyelinating disorders, present significant challenges for accurate diagnosis and personalized treatment due to heterogeneity in lesion distribution, morphology, and progression. Early and accurate lesion identification is critical for disease classification, prognosis, and treatment planning. MRI provides rich tissue-level structural information for lesion detection, while microscopy reveals morphology and topology at the cellular level. Together, they offer complementary perspectives on multifocal brain lesion pathology. However, existing deep learning methods for multimodal MRI segmentation often hinder interpretability and introduce redundant parameters, leaving AI decision-making opaque in clinical settings. My contributions are: 1.SwitchNet: A modality-adaptive fusion network that explicitly learns each MRI modality's contribution to different lesion subregions via Adaptive Encoder/Decoder Blocks and modality-specific attention, improving interpretability while maintaining segmentation accuracy. 2.Multi-scale Diffusion Segmentation: I refactor segmentation into stepwise mask reconstruction, constructing a mask pyramid with independent diffusion processes at each scale to establish structural-level image-mask correspondence. 3.Skeleton-Guided Microscopy Pipeline: A cellular-level analysis pipeline for tumor and non-tumor cell segmentation, topology extraction, and quantitative analysis of cell spatial organization. Includes publication

Inter-project Learning in Major Infrastructure Projects

2026-01-01T00:00:00Z

Inter-project Learning in Major Infrastructure Projects Liu, Yuhan In project-based organizations (PBOs), the episodic character of projects disrupts durable knowledge retention and organizational memory. While inter-project learning (IPL) has been proposed to improve knowledge flow across projects, institutionalizing knowledge does not automatically activate it. Knowledge can remain dormant in manuals, platforms, and reports unless organizations help people reinterpret and adapt it for new project contexts. Drawing on a longitudinal case study of Beijing Capital International Airport Terminal 3 and Beijing Daxing International Airport, led by Capital Airports Holding Company (CAH), this thesis investigates how organizations can sustain and reactivate IPL. The study develops an integrated framework that explains how project knowledge is identified and externalised, embedded through formal systems and routines, and regenerated across later project contexts. It further examines how dormant knowledge is brought back to life through knowledge translation, including deconstructing prior knowledge, reframing it for local use, embedding it into routines and discourse, and renewing it through feedback loops. Rather than treating knowledge as a static object to be transferred, this thesis conceptualises IPL as a reflexive and recursive process grounded in communicative practices and organizational design. By theorising IPL as organizational reproduction and knowledge translation, the study offers insights into how organizations maintain learning continuity and make past project knowledge meaningful and usable again amid fragmentation, turnover, and temporariness. Includes publication

Towards Efficient High Fidelity 3D Reconstruction and Novel View Synthesis

2026-01-01T00:00:00Z

Towards Efficient High Fidelity 3D Reconstruction and Novel View Synthesis Chen, Haodong This thesis studies three-dimensional (3D) reconstruction and novel view synthesis (NVS) under tight hardware and data budgets, a common constraint in VR/AR, robotics, autonomous navigation, and cultural heritage. Existing methods often rely on multi-camera rigs, dense imagery, calibrated cameras, or heavy compute, limiting embedded and dynamic deployment. The thesis addresses this through three largely orthogonal pipeline-stage contributions: E2V for event-stream sensing, 3DLS for splatting-kernel representation, and HDGS for sparse-view depth supervision. First, E2V shows that a single neuromorphic event camera can support full volumetric reconstruction. It uses an end-to-end model to map raw events to voxel occupancy in one pass, without known intrinsics/extrinsics or multi-stage processing. A synthetic dataset of 39,739 object scans generated with an event-camera simulator is also released for training and benchmarking. Second, 3D Linear Splatting (3DLS) improves point-based radiance-field rendering by replacing the Gaussian fall-off in 3D Gaussian Splatting (3DGS) with a bounded linear kernel. This reduces blur, preserves high-frequency detail, and improves rendering speed by 30% while maintaining competitive image quality across benchmarks. Third, Hierarchical Depth-Guided Splatting (HDGS) addresses geometric inconsistency in sparse-view splatting. Its Cascade Pearson Correlation Loss (CPCL) supervises depth across scales, improving geometric accuracy and consistently outperforming prior methods when only a few views are available. Together, E2V, 3DLS, and HDGS advance 3D reconstruction and NVS toward practical use under tight energy, compute, and memory budgets. Remaining challenges include the synthetic-to-real gap, broader kernel design for splat-based rendering, and geometric consistency under extreme sparsity, occlusion, and scene dynamics.

The quantitative significance of membrane lipids to the phosphorus economics of Australian native plants

2026-01-01T00:00:00Z

The quantitative significance of membrane lipids to the phosphorus economics of Australian native plants Liang, Grace Phosphorus (P) is an essential macronutrient that commonly limits the productivity of plants in terrestrial ecosystems worldwide. The prevalence of P-impoverished soils across the Australian continent, and throughout much of Australia’s geological history, has played a significant role in the evolution of its native flora. Many native species exhibit a wide range of traits and mechanisms that allow them to persist and thrive on some of the most P-impoverished soils, including reducing their allocation of P into membrane phospholipids. Membrane phospholipids, which account for around one-third of organic P in leaves, is a unique P compound as they can be substituted by functionally similar non-P lipids, such as galactolipid and sulfolipids. This process of substitution allows for lipidbound P to be liberated and reallocated to maintain plant growth and function. Therefore, phospholipids likely play an important role in the P economics of plants. The main focus of this thesis is on foliar membrane lipids and their quantitative significance to the P economics of native Australian flora. In particular, I investigated the contribution of membrane phospholipids in the adaptation of Australian native plants to P-impoverished soils. Includes publication

Using Genomic Data to Elucidate the Evolution and Historical Biogeography of the Liverwort Family Lepidoziaceae (Jungermanniales; Jungermanniopsida)

2026-01-01T00:00:00Z

Using Genomic Data to Elucidate the Evolution and Historical Biogeography of the Liverwort Family Lepidoziaceae (Jungermanniales; Jungermanniopsida) Rayos, Antonio Jr Luciano Lepidoziaceae, the third-largest family of liverworts, account for almost 10% of liverwort species diversity worldwide, and there are many unanswered questions about the phylogeny and biogeography of this large and diverse family. The broad aims of this thesis were to resolve the phylogenetic relationships among the taxa within Lepidoziaceae, to propose necessary taxonomic revisions based on the resulting phylogenetic analyses, and to provide insights into the spatial and temporal distribution of the family. The thesis opened with an introduction to Lepidoziaceae, the phylogenetic studies that have been done on the family so far, and the potential of genomic data to resolve the phylogeny of the family. The remaining chapters then addressed the aims using a range of methods to analyse data sets with overlapping taxon sampling. Overall, this thesis has presented several interesting findings about the evolutionary relationships among the members of Lepidoziaceae at different taxonomic levels and provided insights into the distribution of the family across space and time. These results have taxonomic implications that can be used as the basis for taxonomic revisions. The findings from this thesis still leave many unanswered questions that can be addressed in future studies.

WOMEN-FRIENDLY QURʾĀN TRANSLATIONS IN COMPARATIVE PERSPECTIVE: A FEMINIST CRITICAL DISCOURSE ANALYSIS OF LALEH BAKHTIAR’S THE SUBLIME QURʾĀN (Q.2:231, Q.4:11, Q.4:34)

2026-01-01T00:00:00Z

WOMEN-FRIENDLY QURʾĀN TRANSLATIONS IN COMPARATIVE PERSPECTIVE: A FEMINIST CRITICAL DISCOURSE ANALYSIS OF LALEH BAKHTIAR’S THE SUBLIME QURʾĀN (Q.2:231, Q.4:11, Q.4:34) Elmir, Mouna This thesis investigates how English translations of the Qurʾān shape gender-conscious interpretations of verses concerning women's rights and roles. It examines Laleh Bakhtiar's The Sublime Qurʾān (2007), focusing on her translation of three Qurʾānic verses (Q.2:231, Q.4:11 and Q.4:34) that have traditionally been associated with debates on marriage, divorce, inheritance and domestic authority. The study argues that Qurʾānic translation plays a significant role in shaping ethical understandings and social attitudes towards women. It explores how alternative translation choices can challenge interpretations that have been used to justify spiritual abuse, domestic violence and patriarchal authority, while remaining grounded in the Qurʾān's ethical framework. Using an interdisciplinary methodology, the study integrates Translation Studies with Critical Discourse Analysis, Feminist Critical Discourse Analysis, Feminist Poststructuralist Discourse Analysis and Contrastive Analysis. Bakhtiar's translation is analysed in comparison with selected classical and modern Qurʾānic commentaries across textual, discursive and contextual levels. The findings demonstrate that Bakhtiar's approach is best understood as an ethically grounded and linguistically informed reinterpretation rather than a feminist ideological project. The thesis makes two original contributions to Qurʾānic translation studies by introducing the analytical category of woman-friendly translators and the concept of woman-friendly commentary as a hybrid exegetical model of tafsīr. Together, these contributions expand contemporary approaches to Qurʾānic translation and interpretation by foregrounding ethical coherence, contextual sensitivity and the Qurʾān's inclusive moral vision.

Young Children’s Meaning-Making through the Use of Tablets in Digital Play

2026-01-01T00:00:00Z

Young Children’s Meaning-Making through the Use of Tablets in Digital Play Guo, Zejian In recent decades, digital technologies have become not only prevalent in early childhood school environments but increasingly embedded in home settings. Digital play has become a significant activity where young children engage, explore, and express themselves. This situation raises important questions about how children make meaning through digital play and how parents, educators, and researchers can understand and support these processes. The current study investigated the meaning-making processes of four young children during digital play at home in Australia and China. The children were accompanied by their parents during these processes. Grounded in sociocultural theories, this study explored digital play behaviour using the digital play framework (DPF) (Bird & Edwards, 2015). The research further examined the roles of oral language and child–parent interactions in shaping these processes, and potential sociocultural differences between two distinct contexts. Based on a qualitative case study methodology, the findings suggested that the four children’s meaning-making processes were informed by their use of oral language, their agency during play, the selection and affordances of technologies, and the child–parent interactions. Various types of parental scaffolding were observed as significant in fostering the children’s communication, creativity, and learning within digital play. The application of the DPF highlighted both epistemic and ludic play, with three distinct digital play movements emerging across the four cases. Cross-case analysis suggested that although cultural differences existed between the Australian and Chinese families, more similarities than differences were observed. It concludes by discussing the broader implications for research and practices for educators, policymakers, parents, and technology designers while acknowledging the study’s limitations and identifying directions for future research in this evolving field.

Leveraging gene-by-environment interactions to identify molecular drivers of diet-induced metabolic disease

2026-01-01T00:00:00Z

Leveraging gene-by-environment interactions to identify molecular drivers of diet-induced metabolic disease Cutler, Harry Benjamin Understanding why individuals differ in susceptibility to cardiometabolic disease remains a central challenge in metabolic research. Although caloric excess is a major driver of obesity and disease risk, the threshold at which dysfunction develops varies substantially between individuals. This points to intrinsic biological mechanisms that buffer against adverse dietary environments. To identify such mechanisms, I developed a systematic experimental pipeline integrating genetic diversity across inbred mouse strains with multi-omic profiling of metabolically relevant tissues following high fat high sugar (HFHS) feeding – a perturbation that induces metabolic disease only in a subset of strains. Comprehensive metabolic phenotyping revealed that tissue-specific dysfunction was strongly shaped by genetic background, in some cases fundamentally altering which tissues were affected by HFHS exposure. Deep learning models trained solely on protein abundance in chow-fed mice accurately predicted strain- and tissue-specific responses to HFHS feeding, leading to the hypothesis that dietary stress responses are governed by the pre-existing molecular architecture of each tissue. To test this, I established a bespoke tissue-specific adeno-associated virus platform to overexpress candidate regulators identified by the deep learning models, substantially accelerating functional validation compared to conventional genetic approaches. This enabled testing of seven candidate regulators of diet-induced metabolic dysfunction. Hepatic SLC22A7 emerged as a previously unrecognised mediator of insulin resistance in both liver and white adipose tissue under HFHS conditions. Collectively, this work demonstrates that integrating systems-level molecular profiling with predictive modelling and high-throughput in vivo validation is a powerful strategy to uncouple metabolic disease from caloric excess, providing a framework for precision therapeutic discovery.

Genetic Architecture and Molecular Mechanisms of Coat Colour Variation in Domestic Dogs

2026-01-01T00:00:00Z

Genetic Architecture and Molecular Mechanisms of Coat Colour Variation in Domestic Dogs Brancalion, Lillian Valmai Coat colour in the domestic dog (Canis lupus familiaris) is a complex trait shaped by several interacting genetic variants, regulatory mechanisms and breed-specific evolutionary history. While multiple pigmentation loci have been identified, the genetic basis of many coat colour phenotypes remains unresolved. This thesis advances knowledge of canine pigmentation genetics by identifying novel genetic variants underlying coat colour variation, resolving complex inheritance patterns, and developing practical frameworks for genetic coat colour prediction. This thesis characterises two variants associated with white-spot modification near USH2A, resolving the molecular basis and inheritance of ticking and roan phenotypes through a three-haplotype allelic series and identifying USH2A as a novel candidate pigmentation gene. Further, this thesis demonstrates that pheomelanin intensity in the Golden Retriever can be accurately predicted using a simple, biologically informed predictive model. Evaluation across multiple dog breeds highlights the importance of population-aware approaches to phenotype prediction. Finally, characterisation of the MITF-A pseudogene and reference genome misassemblies at the MITF locus facilitate more accurate genetic interpretation of this important pigmentation region. Collectively, this research expands understanding of the genetic architecture of canine pigmentation by identifying novel coat colour variants, clarifying complex genomic regions, and demonstrating how biologically informed models can translate complex traits into practical predictive tools. These findings provide valuable Includes publication

Relationship between Metabolome and Survival Characteristics of Rhizobia after Growth in Liquid Media

2026-01-01T00:00:00Z

Relationship between Metabolome and Survival Characteristics of Rhizobia after Growth in Liquid Media Vathshalyan, Nishanthi Legume inoculation enhances biological nitrogen fixation, but desiccation stress reduces rhizobial survival on seed. This thesis investigated the effect of different liquid media on the rhizobial metabolome and desiccation tolerance. The effect of carbon sources and carbon-to-nitrogen (C:N) ratios in defined media were examined and defined medium (mJMM) was compared with complex media, peat extract (PE) and a food-waste-derived medium (FW). Three rhizobial strains representing clover, pea and soybean symbionts were studied. Carbon source strongly influenced metabolome and desiccation tolerance. Growth on L-arabinose increased trehalose levels and improved survival after drying while lower C:N ratios also improved survival. Early survival was positively associated with trehalose, whereas longer-term survival correlated with metabolites including 3-hydroxybutyric acid. To explore mechanisms in complex media, TA1 survival was monitored during storage at 4 °C in PE, FW and mJMM. Although mJMM maintained the highest viable counts, survival was greatest in FW during early stages and in PE after prolonged storage. Pantothenic acid and ribonic acid were positively associated with survival on beads. Despite upregulation of otsB in complex media, intracellular trehalose accumulation was low and did not predict survival. Improved survival was more closely linked to other metabolites and physiological changes. Membrane integrity increased during storage, particularly in PE-grown TA1. Survival correlated positively with saturated fatty acids, xylitol and related metabolites and negatively with unsaturated fatty acids. These findings suggest that improved desiccation tolerance in TA1may be explained by membrane remodelling triggered by growth and storage in PE. Overall, this study advances understanding of the metabolic and physiological basis of rhizobial desiccation tolerance and informs future development of inoculant formulations for seed survival.

Complex Individuality: The Spatial Temporal and Agential Dimensions of The Problem of Biological Individuality.

2026-01-01T00:00:00Z

Complex Individuality: The Spatial Temporal and Agential Dimensions of The Problem of Biological Individuality. Mann, Rebecca This dissertation explores how we understand and demarcate the spatiotemporally unified, cohesive wholes that we call biological individuals. There to be four main points of contention underlying the so-called problem of biological individuality: (1) Whether there is a single kind of biological individual or multiple; (2) How accounts of biological individuality address how different sorts of individuals exist in space and over time; (3) Which biological phenomena or criteria underlie biological individuality; and (4) How different concepts of biological individuality relate to one another and when they come apart. I develop a three-pronged account of biological individuality that recognises three changeable kinds of biological individuals: evolutionary individuals; organisms; and agential individuals. This is based on three core processes in the biological world: evolution, energy, and action. I begin by examining the ambiguity underlying the problem of biological individuality itself, surveying the numerous solutions to the problem. After dispelling monist approaches to biological individuality, I argue that we should treat biological individuality as an umbrella concept, under which there are multiple kinds of biological individuals based on the important ways in which biological processes form unified, cohesive wholes. I then show how the problem of biological individuality has both spatial and temporal dimensions, and that any solution must address both. I then develop a novel metabolic account of the organism, defining the organism as having a centred metabolic network, before arguing for an agential account of biological individuality, which is not coextensive with organismal or evolutionary individuality concepts. This analysis not only sheds new light on many commonly discussed problem cases, such as social insects and aggregative slime moulds, it also illuminates often-overlooked complex cases of biological individuality, including mammals like ourselves. Includes publication

Stand to your post: The impact of incentives on retention in the Australian Defence Force

2026-01-01T00:00:00Z

Stand to your post: The impact of incentives on retention in the Australian Defence Force Plummer, James The Australian Defence Force has publicly stated its strategic objective to expand operational capability. An integral part of that strategy is an increase in the number of service personnel, which requires higher rates of enlistment and (or) retention. This thesis considers three aspects of military service or reward that are expected to influence the retention decisions of personnel; retention bonuses, housing and relocation. It employs data rarely available to researchers, and uses empirical methods to identify how various factors influence the retention of military personnel. This thesis is intended to further academic understanding of a relatively unexplored aspect of labour economics, namely military workforces. It is also intended to inform decision-making by policymakers on a topic of national importance.

Assessment of pathological features to improve accuracy of diagnosis, classification and prognosis of primary and metastatic melanoma

2026-01-01T00:00:00Z

Assessment of pathological features to improve accuracy of diagnosis, classification and prognosis of primary and metastatic melanoma Rawson, Robert V. The gold standard for the diagnosis of melanocytic lesions and provision of vital prognostic information for primary melanoma, is microscopic examination by a trained pathologist. When interpreting difficult to diagnose primary melanocytic lesions the challenges for the reporting pathologist include subjective diagnostic criteria leading to significant interobserver variability in diagnosis, definitive diagnosis when interpreting an initial partial diagnostic biopsy and, more recently, how to integrate the complex genomic information available from ancillary testing with the morphological clues to diagnosis which have been accumulated and learnt over the last 50 or so years by pathologists. Once the diagnosis is made the pathologist’s role is to report key prognostic information to ensure the lesion is appropriately classified, staged and managed. Traditionally pathologists role in the reporting of patients with advanced (macroscopically detectable metastatic melanoma) has been minimal, usually limited to the initial diagnosis and identification of further adverse features including tumour bulk and extranodal extension4. However, the recent development of effective systemic therapies, particularly immune checkpoint inhibitors, have been tested with proven superior survival in the neoadjuvant setting compared with the adjuvant setting5, 6. This provides the opportunity for pathologists to provide crucial information when analysing these neoadjuvant specimens to assist in predicting clinical outcomes in patients and guide further management. The ensuing body of work studies novel ways in which pathological assessment, through morphological and immunohistochemical techniques, of both primary melanocytic lesions and neoadjuvant specimens in metastatic melanoma can lead to more appropriate management and improved patient outcomes. Includes publication

Using Machine Learning to Solve Stochastic Differential Equations

2026-01-01T00:00:00Z

Using Machine Learning to Solve Stochastic Differential Equations Miltchinov, Nikola Dimitrov Scientific machine learning is a rapidly growing field that draws from and combines many disciplines, including computational mathematics, computer science, physics and more. In recent years, great attention has been turned towards a set of methods which allow one to employ machine learning techniques to solve differential equations. These so-called `physics-informed' methods are powerful tools which do not suffer many of the usual limitations of traditional solvers, making them incredibly well-equipped to handle those complex problems which have eluded researchers for years. Up until now, these machine learning models have focused primarily on deterministic differential systems, with little attention paid to those driven by randomness or noise. However, these stochastic problems can be just as interesting, and are by no means less important than those of the deterministic variety. The content presented in this thesis centers around the application of existing machine learning solvers to problems in stochastic calculus, something which has rarely been attempted, and which has up until now excluded a vast range of meaningful problems from consideration. Of particular interest is the Landau-Lifshitz-Gilbert (LLG) equation, which presents unique challenges when attempting to solve. We find that physics-informed solvers are not particularly well suited for dealing with problems of a stochastic nature - at least in their most basic forms - though they do have potential. Whilst there are certain methods which can somewhat bridge this gap, it will ultimately take more research before the models handling stochastic differential equations are able to catch up with those of their deterministic counterparts.

Molecular and Cellular Characterisation of Non-coding Variants in Inherited Heart Disease

2026-01-01T00:00:00Z

Molecular and Cellular Characterisation of Non-coding Variants in Inherited Heart Disease Singer, Emma S. Genetic variants outside of protein-coding regions of genes, i.e non-coding variants, are increasingly identified to cause inherited diseases. However, the clinical relevance of these variants is challenging to determine from sequence context alone, leaving many classified as variants of uncertain significance. This thesis aimed to study the functional consequences and clinical relevance of non-coding variants in people with inherited heart disease and sudden cardiac death (SCD) through genetic and cellular studies. A rare variant burden analysis in Chapter 2, found an enrichment of RNA splice-disrupting variants in six cardiac disease genes in over 1000 people with inherited heart disease or SCD compared to population controls, supporting their relevance to disease causation. RNA extracted from blood amplified two-thirds of cardiac disease genes, highlighting the suitability of blood RNA for functional analyses of splicing variants. Chapter 3 presents functional analysis reports confirming the pathogenicity of eight putative splice-disrupting variants following blood RNA analyses. RNA and protein studies using primary heart tissue samples presented in Chapter 4 confirmed that a 5’ untranslated region deletion in TAFAZZIN abrogated protein expression from the gene, leading to haploinsufficiency and causing Barth syndrome. Eight MYBPC3 midigenes were developed in Chapter 5 to functionally assess the impact of putative splicing variants without requiring patient-derived RNA samples. In Chapter 6, RNA sequencing of patient-specific induced pluripotent stem cell cardiomyocytes confirmed the pathogenicity of a minor intron variant in SCN5A. CRISPR-Cas9 correction of this variant in Chapter 7, reversed cryptic splicing to wild-type. Collectively, data presented in this thesis led to a clinically meaningful reclassification of 13/21 (62%) non-coding variants and demonstrated how incorporating RNA analyses in genetic testing shortens the diagnostic odyssey for families. Includes publication