Browse EPSRC projects you can join [PDF 194KB]

Page 1

Engineering and Informa�cs Project: 2026ENGINF01 Addi�ve manufacturing of s�muli responsive scaffolds for biomedical applica�ons Supervisor: Elizabeth Rendon-Morales Biomaterial based scaffolds have emerged as a poten�al interven�on for personalized drug delivery, wound healing and skin �ssue regenera�on. These are intricate three-dimensional structures, serving as carriers for cells and medica�ons, facilita�ng their integra�on and delivery into the body. This research project in collabora�on with Queen Victoria Hospital is about improving the design and fabrica�on of 3D printed scaffolds for applica�ons in drug delivery and regenera�ve medicine. We will inves�gate the use of addi�ve manufacturing (AM) technologies combined with advanced materials that have the ability to change their shape, proper�es and func�ons over �me in response to external s�muli such as magne�c fields and light sources. It also explores the adop�on of novel robo�c based volumetric imaging and its integra�on with AI tools for the development of personalised 3D printed scaffolds models for �ssue regenera�on. Project: 2026ENGINF02 Design and development of high precision robo�c interfaces and AI-based naviga�on for biomedical applica�ons. Supervisor: Rodrigo Aviles-Espinosa This project is about improving human-assisted high precision biomedical tasks aiming to achieve an increasingly automated approach. It will inves�gate novel approaches and methodologies for improving robot mo�on using AI based-naviga�on. The project will inves�gate the pathways to enhance autonomy in our high precision robo�c systems by adop�ng image-based visual servoing (IBVS), posi�on-based visual servoing (PBVS), and hybrid approaches. This is to be achieved by combining novel AI and machine learning techniques for image quan�fica�on and analysis including deep learning algorithms to automa�cally segment, register, to enable both autonomous naviga�on and high-speed quan�ta�ve diagnos�cs.

Project: 2026ENGINF03 Scene and movement alignment toward human mo�on genera�on Supervisor: Temitayo Olugbade Scene-aware genera�on of synthe�c human data requires aligned scene and human movement data whereas a large number of exis�ng human movement datasets are without corresponding scene data. This project will inves�gate alignment of related 3D scenes and human movement data from separate datasets. The project will use exis�ng scene and human movement datasets. It will specifically compare a machine learning approach with a constraint-based method. Sui et al. (2026). A survey on human interac�on mo�on genera�on. Interna�onal Journal of Computer Vision, 134(3), 113.

Page 2

Wang et al. (2022). Humanise: Language-condi�oned human mo�on genera�on in 3d scenes. Advances in Neural Informa�on Processing Systems, 35, 14959-14971. Olugbade et al. (2022). Human movement datasets: An interdisciplinary scoping review. ACM Compu�ng Surveys, 55(6), 1-29. Olugbade et al. (2024). The EmoPain@Home dataset: Capturing pain level and ac�vity recogni�on for people with chronic pain in their homes. IEEE Transac�ons on Affec�ve Compu�ng. Dai et al. (2017). ScanNet: Richly-annotated 3D Reconstruc�ons of Indoor Scenes. Proc. Computer Vision and Patern Recogni�on (CVPR).

Project: 2026ENGINF04 Cardiac electrogram classifica�on using deep learning methods Supervisor: Luc Berthouze This project focuses on the automated classifica�on of atrial fibrilla�on (AF) electrograms. The dataset consists of endocardial and epicardial electrograms recorded under controlled pacing protocols from two pa�ent groups: (i) pa�ents with clinically diagnosed AF, including paroxysmal, persistent, and long-standing persistent forms; and (ii) pa�ents without AF who undergo invasive cardiac mapping for other cardiac condi�ons. All recordings are obtained during pacing to ensure clean, repeatable, and comparable electrogram data across pa�ents and condi�ons. By working exclusively with paced AF electrograms, the project targets differences linked to underlying atrial substrate and AF phenotype rather than variability arising from spontaneous rhythm dynamics. The work is embedded within an ongoing collabora�on with University Hospital Sussex NHS Founda�on Trust (UHSx) and contributes to a broader programme developing data-driven methods for analysing atrial electrophysiology in clinically realis�c se�ngs.

Project: 2026ENGINF05 Can Neural Networks Recover Meaning from Encrypted Text? Exploring the Geometry of Encryp�on in Embedding Spaces Supervisor: Olayinka Ajayi This project explores a fundamental ques�on at the intersec�on of machine learning and cryptography: Can a neural network recover the seman�c meaning of encrypted text without explicitly performing classical decryp�on? In classical cryptography, plaintext is transformed into unintelligible ciphertext, and decryp�on restores the original text exactly. In contrast, modern deep learning models are capable of extrac�ng seman�c structure from noisy or corrupted inputs. This project inves�gates whether seman�c informa�on survives encryp�on in a form that is detectable in high-dimensional embedding spaces

Page 3

Project: 2026ENGINF06 Title: AI-Assisted Op�cal Diffrac�ve Models for Real-�me Image Processing Supervisor: Mohammad Neshat This project inves�gates AI-assisted op�cal diffrac�ve models for real-�me image processing using simplified computa�onal simula�ons. The aim is to explore how op�cal-inspired systems— specifically diffrac�ve layers—can process visual informa�on at the speed of light, offering a poten�al alterna�ve to conven�onal digital compu�ng. Such approaches are highly relevant to emerging applica�ons where ultra-fast and energy-efficient processing is cri�cal. Examples include real-�me object recogni�on in autonomous vehicles, high- speed quality inspec�on in manufacturing, biomedical imaging for rapid disease detec�on, and next- genera�on augmented/virtual reality systems requiring low-latency vision processing. Op�cal diffrac�ve compu�ng also has poten�al in security and sensing, enabling instant patern recogni�on without heavy digital hardware. In this project, the student will develop a 2D numerical model where images are treated as op�cal fields that propagate through a small number of virtual diffrac�ve layers. These layers, represented by trainable masks, will be op�mized using basic machine learning techniques to perform a simple task such as digit or patern classifica�on. The model will be evaluated under ideal and noisy condi�ons and compared with a standard digital approach.

Project: 2026ENGINF07 Title: Opportuni�es and Challenges of Automa�ng Topic Analysis using LLMs Supervisor: Maria Teresa Llano Topic analysis is a qualita�ve research method widely used in the research community; however, it requires a lot of resources, such as �me and people. Current AI technologies, such as cloud-based LLMs, have been used to automate parts of the process of topic analysis in different domains. The idea of this project is to develop a proof of concept system that automates stages of topic analysis using local LLMs instead of cloud-based LLMs. An exis�ng survey paper will serve as the baseline for comparison of performance between a manual and an automa�c approach through the system.

Project: 2026ENGINF08 Electricity from underground: UK geothermal power Supervisor: Mar�n White It is an exci�ng �me for geothermal power in the UK. At the beginning of March 2026, the first geothermal power plant in the UK was switched on. The plant will use heat extracted from deep underground to generate electricity for over 10,000 homes. Geothermal power is an excellent addi�on to our na�onal energy mix, providing clean, base-load energy that is available 24-hours, and 365-days a year. Geothermal power could meet a fi�h of global electricity needs, but amongst challenges related to long development �melines, and challenging policy and regulatory frameworks, high costs remain a key roadblock.

Page 4

This project will carry out the techno-economic op�misa�on of an innova�ve concept for surface- level geothermal power plants that could improve power output by up to 30%, thereby enabling a step change in the revenue from a given geothermal reservoir over its lifespan to help overcome financial barriers. The concept is based on the idea of wet-to-dry expansion, where molecularly complex fluids are exploited inside the power plant to enable the efficient expansion of liquid-vapour mixtures to fully vapour condi�ons using a turbine. The aim of this project will be to develop a techno-economic model of the power plant that can assess both the thermodynamic performance of the plant, as well as the capital cost of key components such as the heat exchangers, pump and turbine. This model will be coupled to a suitable op�misa�on rou�ne to iden�fy op�mal system designs for different geothermal applica�ons, and these designs will be benchmarked against conven�onal geothermal power plants to assess the poten�al benefit.

Project: 2026ENGINF09 Supervisor: Mar�n Berger I support various projects in this space. I am open-minded regarding the exact details of the project. Please contact me (M.F.Berger@sussex.ac.uk) for ques�ons and to discuss ideas, below are examples of projects I would be interested to support. Implemen�ng automa�c theorem proving on GPUs - GPUs are the work-horses of high-performance programming. The accelera�on they provide to applica�ons compa�ble with their programming paradigm can surpass CPU performance by several orders of magnitude, as notably evidenced by the advancements in deep learning. A significant spectrum of applica�ons, especially within automated reasoning like SAT/SMT solvers and model checkers has yet to reap the benefits of GPU accelera�on.One conjecture why is that the algorithms used in logic based verifica�on are intrinsically sequen�al and cannot be parallelised. I conjecture that that is false. An interes�ng project space would be to experiment with the implementa�on of the corner stone of all logic, namely decision procedures. It would be fun to implement at least one such decision procedures on a GPU and see how fast they can be made. ARC challenge - The ARC challenge (short for: Abstrac�on and Reasoning Challenge) is one of the most important AI benchmarks. It was designed specifically to be difficult for deep-learning. It is intended to measure how close we are to AGI. They mostly use either program synthesis or LLMs, but mostly combine both. I think developing interes�ng ideas to do with program synthesis and program search (with or without LLMs) is especially fun. Computer security at the hardware / so�ware boundary - More and more security vulnerabili�es, happen at the hardware / so�ware boundary, examples include Spectre, Meltdown, and Rowhammer. Formalising mathema�cs or theore�c computer science in Lean (or similar) - Interac�ve theorem provers like Lean (htps://en.wikipedia.org/wiki/Lean(proofassistant)) have recently become popular, both in: deep learning for RLVR (reinforcement learning with verified reward), smashing various mathema�cs related benchmarks (e.g. the IMO Grand Challenge). But also for building formally verified bases for Mathema�cs, High energy physics and Computer science. I'm especially interested in the mathema�cs and computer science of Monte Carlo simula�on, as o�en used in high-energy physics. Just-in-Time (JIT) compila�on with PyPy - Designing programming languages is great fun. Wri�ng an interpreter for a language is easy, but does not deliver speed. Making a new language fast is difficult because performant compilers are expensive to engineer.

Page 5

The problem is especially significant for dynamically typed languages like Python or Javascript, that are intrinsically slow. JIT compilers (short for just-in-�me) make their performance (somewhat) bearable, but at the cost of even more complex compilers. Meta-tracing JITs convert a language interpreter into a JIT compiler, thus removing much of the cost of wri�ng JITs by hand, see my compiler notes for a superficial overview. The PyPy framework is a widely used meta-tracing framework. Programming language design (and implementa�on) - Designing programming languages is great fun. Only our imagina�on is the limit. Designing a dream language is also a great way of becoming a beter programmer. With agen�c-coding much of the grind-work can now be automated away. I think developing interes�ng ideas to do with typing systems and/or meta-programming is especially fun. RISC-V - RISC-V is an open-source instruc�on-set architecture (ISA) designed to be modular. This open nature facilitates academic research and industrial adop�on. Projects include: Wri�ng a RISC-V simulator, Implemen�ng a RISC-V using Verilog, or similar languages, Designing a new RISC-V extension. Semirings and systolic arrays: systolic arrays for Semiring - Semirings are one of the most general and widely used mathema�cal structures. They generalise rings, by weakening the requirement that addi�on is an abelian group. In a semiring, addi�on needs only be a commuta�ve monoid. Matrix mul�plica�on can be defined for general semiring. In recent years TPUs and tensor cores came to the for as hardware accelerators for deep learning. They are typically all based on systolic arrays a simple processor architecture that is especially good for matrix opera�on like matrix mul�plica�on. To date all those TPUs and tensor cores hard-code floa�ng point numbers (which makes sense for neural networks). I think it would be fun, to generalise this a �ny bit and build a systolic arrays for general semirings. AI-agents as research helpers - The last few months have seen an explosion in the ability of LLMs to aid science. This is typically driven by some feedback loops, including RLVR (reinforcement learning with verified rewards) in training. O�en this is combined with feedback loops involving users (e.g. Agent-Driven Development in programming). Some�mes it is also combined with other search techniques. Pydrofoil (or QEMU) for processor simula�on - Before processors are "taped out" meaning: printed to silicon, they undergo extensive design and verifica�on stages. But since, in those stages, they don't exist as physical processor, they are designed and verified using simula�on. One well-known such simulator is QEMU that is widely used in industry. I have been involved with Pydrofoil based on PyPy.

Research areas