CSC PhD Studentships in Electronic Engineering and Computer Science

Supervisor: Dr Ahmed M. A. Sayed

Large Language Models (LLMs) deliver advanced reasoning capabilities but remain computationally prohibitive for distributed and embedded environments [1] [2]. This research will investigate resource-efficient inference in networked AI systems, where multiple edge/embedded devices collaboratively host and execute LLM components. It will develop communication-aware partitioning, KV cache optimization, and dynamic workload scheduling to minimize latency, memory footprint, and inter-node bandwidth. Through hardware–software co-design [3] [4], the work will align LLM execution with accelerator hierarchies and network topology for efficient distributed inference. The outcome will be an architectural framework and scheduling algorithms enabling scalable, energy-efficient, and cooperative LLM deployment across interconnected edge and IoT environments.

Primary supervisor: Akram Alomainy 

Second supervisor: Mohamed Thaha

Wireless capsule endoscopy (WCE) is an innovative medical technique used for diagnosing and treating gastrointestinal tract conditions. Conventional endoscopic procedures face challenges in adequately scanning the small intestine due to its anatomy and location. Currently, capsule endoscopy is primarily utilized as a diagnostic tool but has certain limitations. As a result, the development of future capsule prototypes seems unavoidable. Advancements in WCE technology aim to overcome these limitations. These advancements include three-dimensional reconstruction of high-resolution images, high-frame-rate imaging, complete spherical imaging, and capsule chromoendoscopy. By employing these techniques, unnecessary invasive examinations can be minimized as the endoscopic and microscopic features of small intestinal lesions can be clearly visualized. This offers improved diagnostic capabilities and potentially reduces the need for more invasive procedures. The primary objective of this project is to develop miniaturized antennas with high-data-rate capabilities for wireless capsule endoscopy, utilizing innovative techniques. Additionally, the project aims to perform a clinical evaluation of current wireless capsule endoscopy systems to create appropriate testing environments.

Primary supervisor: Dr Anna Xambó Sedó

Second supervisor: Dr Charalampos Saitis 

Large language models (LLMs) are a type of artificial intelligence program that can recognise and generate text, which are trained on huge sets of data with a complex network of hidden processes. This PhD topic explores sonification techniques of LLMs for a better understanding of the way they process the information. Can we treat LLM engines such as ChatGPT as a musical instrument and listen to its internal processes? Can sonification techniques help us to hear and see how the information is processed? Compared to vinyl records or tape recordings, what is the acoustic signature, and what are the artefacts that are distinctive of this new medium? This work will contribute to addressing an important challenge in AI: making the inner workings and hidden knowledge of models more interpretable for people.

Keywords: sonification, large language models (LLMs), explainable A

Primary supervisor: Anthony Constantinou 

Machine learning has achieved major success in modelling complex patterns from large datasets, yet most approaches emphasise predictive accuracy over interpretability or causal understanding. This limits their ability to support reasoning about hypothetical actions and underlying decision-making mechanisms in real-world systems.

Causal inference and causal discovery methods address this gap by identifying cause-and-effect relationships and providing a principled foundation for interpretability. However, they often rely on strong assumptions, face scalability challenges, and are difficult to apply in heterogeneous or high-dimensional settings.

This project will advance statistical, probabilistic, and causal learning methods while drawing selectively on ideas from deep learning and related advances, such as representation learning, variational inference, and graph-based neural architectures, that have proven effective in other areas of AI. The goal is to adapt and integrate these concepts to improve the scalability, robustness, and flexibility of causal and probabilistic modelling, leading to improvements in causal representation models learning.

Supervisor: Arkaitz Zubiaga 

Large Language Models (LLMs) have led to unprecedented advancements in content production, but have also given rise to adversarial use. This includes generation of deceitful LLM-generated content (e.g. disinformation, fake reviews, fake job applications), harmful content (e.g. LLM chatbots that use offensive and derogatory language) and misleading content (e.g. influencer marketing), among others. Existing methods to deal with this adversarial use of LLMs are limited and ineffective. Moreover, content produced by adversaries varies across languages and across cultures, which makes their detection and mitigation even more challenging in a generalisable manner. Some of the issues arise from how LLMs are designed and calls for improved approaches to safeguarding LLMs, whereas other issues arise from how these LLMs are used by malicious actors, which calls for prevention and mitigation of its effects.

This project definition is broad in scope and allows narrowing it down to the specific interests of the applicant, and hence you are welcome to reach out to me for further discussion of your interests and to discuss specific directions.

Supervisor: Dr Athen Ma

Our natural environments are subjected to reoccurrence of extreme climatic events such as heatwaves, posting serious threats to biodiversity, and ecosystem services, such as pollination and food production. A better understanding on how ecological communities respond and mitigate the effects of an environmental stressor will therefore greatly enhance our ability to guide conservation and management efforts.

Species do not exist in isolation but is part of a complex organisation of interactions, manifesting into non-random structural patterns at the whole-network level. While these structures help explain ecosystem function, dynamics and stability, the principles that underpin their assembly and re-assembly are largely unknown. In network science, theoretical concepts based on optimisation of cost or reachability have been used to characterise the way in which complex networks evolve over time, but this line of research is largely unexplored in natural ecosystems. The overall aim of this project is to perform novel merging of network science with community ecology so as to reveal the mechanics that underpin the changes in the topology of ecological networks under climate change. Findings will provide important insights into their temporal dynamics and compensatory response to an environmental disturbance, paving the way for a more predictive approach in ecology.

Supervisor: Changjae Oh 

Vision-Language-Action (VLA) models are designed to bridge the gap between perception and action by leveraging large-scale Vision-Language Models (VLMs) pre-trained on robotic data. However, such models are typically large and computationally heavy, and hence they are difficult to be quickly adapted to new tasks or environments. This project will investigate how such VLA models can be compressed into smaller models that can be quickly adapted to new manipulation tasks while maintaining the perception capability of the large VLM. Specifically, the goal is to develop a lightweight model and learning approach that can reduce reliance on large foundation models while effectively handling manipulation tasks in real-world environments. By the end of the project, the student will have gained a solid background and skills in computer vision, robot learning and manipulation.

Supervisor: Emmanouil Benetos

The field of music information retrieval (MIR) has been growing for more than 20 years, with re-cent advances in deep learning having revolutionised the way machines can make sense of music data. At the same time, research in the field is still constrained by laborious tasks involving data preparation, feature extraction, model selection, architecture optimisation, hyperparameter optimisa-tion, and transfer learning, to name but a few. Some of the model and experimental design choices made by MIR researchers also reflect their own biases.

Inspired by recent developments in machine learning and automation, this PhD project will investi-gate and develop automated machine learning methods which can be applied at any stage in the MIR pipeline as to build music understanding models ready for deployment across a wide range of tasks. This project will also compare the automated decisions made on every step in the MIR pipe-line, as compared with manual model design choices made by researchers. The successful candidate will investigate, propose and develop novel deep learning methods for automating music under-standing, resulting in models that can accelerate MIR research and contribute to the democratisation of AI

Supervisor: Dr Evangelia Kyrimi

In modern healthcare, machine learning and deep learning models are increasingly used to support clinical decision-making. However, their complexity often makes it difficult for clinicians to understand how predictions are derived, limiting trust and adoption. Providing explanations that are both interpretable and causally meaningful is therefore a critical challenge.

We invite applications for a PhD position focused on developing causally informed explanations for ML and DL models in healthcare. This research aims to enhance clinical decision-making by generating explanations that reflect both causal relationships and statistical patterns, while accommodating heterogeneous input data such as structured lab results, time-series measurements, and imaging-derived features.

The candidate will engage in the following areas:

· Develop Algorithms: Transfer lessons learned from explaining causal Bayesian Network models to generate transparent, actionable explanations for more complex ML and DL models.

· Create User-Centric Outputs: Present explanations tailored to the needs of clinicians and researchers to improve understanding and trust.

· Establish Evaluation Metrics: Develop frameworks to assess the interpretability, relevance, and reliability of explanations in clinical scenarios.

By advancing methods for causally meaningful explainability in complex predictive models, this project will support clinicians in making safer and more informed decisions, contributing to the broader adoption of AI in healthcare.

Supervisor: George Fazekas 

Audio and music representation learning seeks to transform raw data into latent representations for downstream tasks such as classification, recommendation, retrieval and generation. While recent advances in deep learning, especially contrastive, self-supervised and diffusion-based approaches have achieved impressive results, most remain purely data-driven and neglect domain-specific musical structures like rhythm, melody, harmony, metrical hierarchy or genre-style traits.

This PhD project will explore ways to embed theoretical and structural knowledge into modern representation learning pipelines to enhance interpretability, controllability and performance. For example, incorporating symbolic or other structured representations, inductive biases, well-known principles exploited in classic DSP algorithms, or ontological constraints, the research aims to bridge the gap between data-driven models and the structured understanding of music and audio.

Potential directions include hybrid models that combine deep audio and symbolic embeddings, graph-based or relational learning of musical structure, and explainable methods for music analysis, production or generation. The project will also engage with principles of Ethical and Responsible AI: reducing data bias, improving transparency and supporting fair attribution of authorship.

Examples of relevant works include but not limited to:

Guinot, Quinton, Fazekas: “Semi-Supervised Contrastive Learning of Musical Representations”, ISMIR-2024

Yu, Fazekas: “Singing voice synthesis using differentiable LPC and glottal-flow-inspired wavetables”, ISMIR-2023

Agarwal, Wang, Richard: F-StrIPE: Fast Structure-Informed Positional Encoding for Symbolic Music Generation, ICASSP-2025

Supervisor: Dr Jin Zhang 

Second supervisor: Prof Xiaodong Chen 

Particle accelerators play an increasingly important role in medicine, environmental protection, industry, and fundamental science. However, their operation demands substantial electrical power, and as the number of accelerators exceeds 40,000 worldwide, their collective energy use has become a critical sustainability concern. For example, the European Organization for Nuclear Research (known as CERN) currently uses 10% of the electricity in the Geneva region. In the UK and worldwide, reducing the environmental impact of accelerator facilities is essential to maintaining scientific competitiveness and meeting the energy-efficiency goals.

This project investigates the development of energy-efficient particle accelerators using high-efficiency radio frequency (RF) power sources. The research will examine how next-generation RF technologies can deliver stable, high-power electromagnetic fields with reduced energy demand, enabling compact, cost-effective accelerator designs for clinical and research use. Emphasis will be placed on adapting high-efficient RF sources, e.g. magnetron, improving phase and frequency control, and enhancing compatibility with superconducting RF (SRF) accelerator architectures. Through simulation and experimental validation, the project aims to demonstrate measurable reductions in energy consumption while maintaining accelerator performance. The outcomes will contribute to the advancement of sustainable medical accelerator technologies, supporting transition towards greener healthcare and energy-efficient scientific infrastructure.

Primary supervisor: Marcus Pearce 

Second supervisor: Iran Roman 

Evidence suggests that speech and music perception depend on cognitive models acquired through cultural exposure via a process of implicit statistical learning. Predictions generated from these models enable efficient and effective processing of rapidly time-varying audio signals leading to culturally appropriate interpretation of meaning. Existing work has successfully simulated these processes of learning and prediction using structured probabilistic models. Meanwhile, deep autoregressive neural network architectures incorporating self-attention have used analogous mechanisms of statistical learning and prediction for convincing generation of language and music. However, it is not known whether these models also simulate perception of speech and music. The proposed project will develop neural network architectures for simulating speech and music perception using existing probabilistic methods both as a benchmark and as a tool for interpreting abstract representations learned by the neural networks. The models will be tested through iterative comparison with behavioural and neural data from psychological experiments with humans. Cross-cultural comparisons of developmental trajectories to adult performance will assess the models as computational simulations of human cultural learning alongside detailed comparisons of the psychological relationships between speech and music. The outcome will be a comprehensive computational understanding of the psychology of human cultural learning in auditory perception.

Supervisor: Arumugam Nallanathan 

A new concept in 6G, known as Integrated Sensing and Communications (ISAC), offers a promising solution for simultaneously achieving radar sensing and wireless communications. The significance of ISAC is reflected in its inclusion as an active work item in major telecommunications standardization bodies, such as 3GPP, IEEE, ITU, Next G Alliance, IMT-2023, and the 6G framework established by ITU-R in June 2023. Despite its potential, ISAC faces numerous challenges in the context of autonomous driving, especially in offering simultaneous sensing and communication with fast response and high accuracy. This project aims to address key issues in ISAC for autonomous vehicles, focusing particularly on interference management, antenna diversity order enhancement, develop novel waveform and signal processing techniques for ISAC and integrate the machine/deep learning based channel estimation with ISA

Primary supervisor: Qianni Zhang 

Second supervisor: Greg Slabaugh

3D vessel registration is a necessary step to understand the implications of blood flow on coronary atherosclerosis. Furthermore, 3D vessel models allow for full visualisation of vessel geometry, stenosis (narrowing) and can assist in stent placement. To assess coronary atherosclerosis, it is often required to employ multiple intravascular imaging techniques, such as intravascular ultrasound (IVUS) and Quantitative Coronary Angiography (QCA). This research will focus on solving the crucial problem of multimodal registration of vessel imaging, and prediction of a 3D vessel model with accurate features, followed by potential modelling and assessment of plaque composition. The integration of these diverse imaging modalities in a detailed 3D vessel model will be further used for implementing fluid simulation to assist the assessment. Barts Hospital has provided comprehensive multimodal vessel imaging data. Experiment results will be evaluated by cardiovascular experts

Primary supervisor: Dr Riccardo Degl’Innocenti

Second supervisor: Prof. Akram Alomainy

Research in the terahertz (THz) frequency range has unlocked unique applications in several areas. Amongst the most important ones, biomedical terahertz imaging is recognized as one of the most promising techniques in diagnostics [1,2]. This is due to the non-ionizing nature of this radiation, allowing for a non-harmful investigation, and to its high sensitivity to the hydrogen bonds and therefore to the hydration level in biological samples, useful to discriminate, for instance, skin cancer tissues from healthy ones [3,4].

3D human skin equivalents (HSE) [5-7], artificially reconstructed epidermal and dermal layers, imitate natural skin and are some of the most robust tissue models currently available. They offer a unique controlled environment in which the composition and structure of the tissue can be precisely manipulated, thus providing an ideal test bed for the development, optimization, and validation of THz imaging. The proposed project will focus on HSEs, provided in collaboration with the Centre for Cell Biology and Cutaneous Research of Blizard Institute, to further progress THz biomedical imaging. The Antenna&EM group has unique infrastructure and facilities, e.g. VNAs, 3 different THz-TDS systems, as well as a previously acquainted expertise in THz spectroscopy of biological materials.

1 J. H. Son, Ed. Terahertz Biomedical Science & Technology, Boca Raton, USA, CRC press, Taylor & Francis Group, (2014).

2 1. Q. Sun et al “Recent advances in terahertz technology for biomedical applications” Quant Imaging Med Surg. Jun;7(3):345-355 (2017).

3 J. Wang et al. “THz Sensing of Human Skin: A Review of Skin Modeling Approaches” Sensors, 21, 3624 (2021)

4 S. Nourinovin et al. "Terahertz Dielectric Characterization of Three-Dimensional Organotypic Treated Basal Cell Carcinoma and Corresponding Double Debye Model," in IEEE Trans. Terahertz Sci, 13(3), 246-253 (2023),

5 S. Nourinovin et al. "Terahertz Characterization of Ordinary and Aggressive Types of Oral Squamous Cell Carcinoma as a Function of Cancer Stage and Treatment Efficiency," IEEE Trans. Instrum. Meas. 72,1-9 (2023).

6 Ahn, M. et al. “3D biofabrication of diseased human skin models in vitro.” Biomater Res 27, 80 (2023).

7 C. A Brohem et al. “Artificial skin in perspective: concepts and applications” Pigment Cell Melanoma Res. 24; 35–50 (2010).

Primary supervisor: Dr SaeJune Park 

2nd supervisor: Dr Riccardo Degl’Innocenti

Terahertz (THz) waves have been intensively explored for the last couple of decades owing to their unique properties e.g., label-free detection of target materials using their spectral fingerprints in the THz frequency range. However, examining target materials with THz waves becomes challenging when the volume of the target material is small compared to the THz wavelength due to the small scattering cross-section between the THz waves and samples. Developing novel platforms to enhance THz spectral fingerprints for detecting extremely small amounts of target materials is imperative to deliver rapid identification of small amounts of target materials for various purposes. This PhD project will study THz devices and systems such as metamaterials and waveguides to enhance the THz fingerprints of the target materials.

Primary supervisor: Simon Lucas 

Second supervisor: Raluca Gaina 

This PhD will explore the use of LLM-generated Code World Models for statistical forward plan-ning algorithms such as Monte Carlo Tree Search and Rolling Horizon Evolution. Large Language Models (LLMs) will be used to generate, test, and refine executable simulation code that captures causal structure and supports the reasoning processes of these algorithms. The student will develop LLM-in-the-loop workflows for automatic code synthesis, debugging, and refinement, optimising both decision performance and predictive accuracy within the chosen problem domains, while maintaining high software quality.

The research will combine methods from evolutionary computation, reinforcement learning, LLMs, and software engineering to construct fast, copyable forward models for complex, dy-namic environments such as games, logistics, or air-traffic control. Evaluation will employ open benchmarks from these domains, assessing efficiency, reliability, consistency, and interpretabil-ity.

This work aims to advance the frontier of agentic AI, developing systems that not only use simula-tions but can write and improve them autonomously—a step towards AI that learns by designing its own models of the world.