|
Safety-Aligned Agentic AI for Clinical Decision SupportPrinciple Investigator: Mobarak HoqueResearch profile Project DescriptionArtificial intelligence has the potential to support clinicians by bringing together information from medical images, clinical notes and other patient data. However, these systems must be able to recognise when they may be wrong and avoid giving overconfident or unsafe recommendations. This project will develop safety-aligned agentic AI systems that can reason across different types of clinical information, assess the reliability of their own outputs and identify when human review is needed. The research will study how these systems behave across different datasets, clinical scenarios and challenging cases, including situations where the available information changes or is incomplete. The goal is to develop accurate, transparent, and reliable AI for real-world clinical decision support.
|
|
Solve Wind Dispersion at Street-Canyon ResolutionPrinciple Investigator: Alessandro De RosisResearch profile Project DescriptionThis project will develop advanced computer simulations to understand how wind moves through urban streets and how pollutants such as traffic emissions spread within cities. By modelling individual streets and buildings in detail, we will capture important airflow patterns that are often missed by lower-resolution approaches. The simulations will use powerful GPU computing facilities, allowing us to analyse realistic urban areas at much higher resolution and over much shorter timescales than would otherwise be possible. The research will help identify how street layout, building geometry and wind conditions influence pollutant concentrations and dispersion. Ultimately, the project aims to provide new computational tools and scientific understanding that can support better urban planning.
|
|
Sustainable AI Data Compression Study: Profiling BOA on BEDEPrinciple Investigator: Caterina DoglioniResearch profile Project DescriptionThis project, supported by the N8CIR internships, will create a training resource in the form of an executable workflow that quantifies the energy consumption of an advanced compression technique for high energy physics data that requires large-scale AI training to be run on a High-Performance Computing (HPC) system such as Bede. The compression algorithm, called BOA Constrictor (published in Machine Learning in Science and Technology, DOI 10.1088/2632-2153/ae64a9) is a Mamba-network based lossless data compressor, and an open source project led by two Early Stage Researchers (PhDs at the University of Manchester) and co-supervised by the academic supervisor of this project and by a RSE at the University of Manchester.
|
|
Accurate Machine Learning Potentials with AutoencodersPrinciple Investigator: Richard BryceResearch profile Project DescriptionThe motivation of this project is to develop new tools for drug discovery and optimisation –specifically, a deep learning model capable of predicting free energy and related properties of drug molecules across conformations for integration into molecular dynamics simulations and virtual screening. In parallel, we will construct a large-scale, chemically diverse dataset of drug-like molecules and conformations with quantum-mechanical energies and interatomic forces. The expected outputs include a transferable machine learning potential, a scalable inference pipeline for arbitrary molecular systems, and a high-quality reference dataset. Together, these resources will support large-scale molecular simulations and future studies on foundation models for drug discovery.
|
|
AI-ExWave: Machine Learning for Extreme Ocean WavesPrinciple Investigator: Tim TangResearch profile Project DescriptionExtreme ocean waves can damage ships, offshore wind turbines and coastal infrastructure. They are difficult to predict because they are rare, develop rapidly and arise from complex interactions between wind and waves. Detailed computer models can represent these processes, but they are too slow to examine the many storms and sea states needed for reliable risk assessment. This project will develop and train a machine-learning model to represent extreme waves more efficiently. We will use established simulations to teach the model wave behaviour during storms, and observations to evaluate its predictions. Bede will provide the computing power to train the model using large datasets and evaluate it across many conditions. We aim to improve wave prediction for offshore and coastal safety.
|
|
Scalable Curvature Methods for RL Policy OptimizationPrinciple Investigator: Yihe ZhouProject DescriptionThis project investigates how to train AI systems more efficiently and reliably. Modern reinforcement learning methods, such as those used to train decision-making agents and language models, often require large amounts of computation and can be unstable. It will use N8 Bede to test new optimization methods that make better use of curvature information during training. These methods aim to improve learning speed, stability, and compute efficiency compared with standard first-order approaches. The project will run controlled experiments on reinforcement-learning tasks and analyse how different optimization methods affect policy improvement and training behaviour.
|
|
Manifold Galerkin Learning for Turbulent FlowsPrinciple Investigator: Miguel BeneitezResearch profile Project DescriptionUnderstanding and controlling turbulence is vital to many physical systems, from large scales affecting the climate to the small scales governing fluids in the human body. In this project, we will develop low-order models of complex turbulent flows to allow for many-query applications, as well as fast design. Obtaining these models will further aid us to unravel the underlying mechanisms of turbulence, when large-scale systems remain largely of reach to be effectively computed or experimented on.
|
|
Expectation-Maximization Methods for Policy OptimizationPrinciple Investigator: Adrian Mircea NenuResearch profile Project DescriptionExploring the use of expectation maximization (EM) for policy optimization in deep reinforcement learning across tasks such as control, large language models and world models. We are identifying adaptations of the EM loop that make this approach viable, and we conduct an ablation study for code-level optimisations.
|
|
Generative Diffusion Model for Air Quality ForecastingPrinciple Investigator: Congbo SongResearch profile Project DescriptionAir pollution remains a critical global challenge, impacting millions of lives and demanding more accurate, timely forecasts. This project leverages cutting-edge Artificial Intelligence—specifically Generative Diffusion Models—to predict air quality with unprecedented precision. By training NVIDIA’s advanced StormCast model on complex atmospheric simulations (WRF-EMEP), we aim to create a "digital twin" of regional/urban air quality. Our research incorporates real-world ground observations into these AI models to forecast air quality levels more reliably. This work seeks to provide clearer early warnings for pollution events, empowering policymakers and the public to make informed decisions for a healthier, more breathable future.
|
|
Understanding Boundaries and Interpretability in LLMsPrinciple Investigator: Chenghua LinPhD Researcher: Yang Wang Student research profile Project DescriptionLarge language models (LLMs) such as ChatGPT are increasingly used in everyday tools, research, and decision-making systems. However, we still have a limited understanding of how these models store information or why they sometimes produce incorrect or misleading answers. This project investigates the knowledge boundaries of LLMs. We also develop new techniques to make these systems more interpretable, helping researchers identify the internal processes that lead to errors or unexpected behaviour. By improving our understanding of how LLMs think and fail, this work aims to support the development of safer, more transparent AI systems that can be reliably used in science, education, and wider society.
|
|
Evaluation Methods for Next-Generation NLG SystemsPrinciple Investigator: Chenghua LinPhD Researcher: Yiqi Liu Student research profile Project DescriptionLarge language models are increasingly used in everyday technologies, including chatbots, online assistants, and text generation tools.However, evaluating how well these systems actually perform remains a significant challenge. My project aims to develop clearer and more reliable ways to assess their behaviour, especially in situations where models may give overly confident or biased answers. By studying how different models respond across a wide range of tasks and datasets, I hope to build evaluation methods that better reflect real world use. The results will help create safer and more trustworthy AI systems for the public.
|
|
Dissecting carbohydrate-protein interactions for drug designPrinciple Investigator: Richard BryceResearch Profile Project DescriptionCarbohydrates play a key role in biomolecular recognition, for processes including cell-cell signalling and viral adhesion, and invasion. The ability to accurately dissect the key regions of interaction between a carbohydrate and its protein receptor is key to understanding its mechanism of action and in guiding the design of drug molecules that could modulate this process. This project seeks to develop a computational strategy to reliably identify the key parts of a carbohydrate that determine its protein-binding ability. The approach is based on calculating free energies using molecular dynamics simulations and will be initially calibrated using reference carbohydrate-lectin datasets before being applied to probe viral oligosaccharide-protein interactions for drug design.Publications
|
|
Integrating Machine Learning into the Finite Element MethodPrinciple Investigator: Lee MargettsResearch Profile Project DescriptionAll products being developed today need to be re-imagined. This is to address increasing demands from the consumer and to conform to tighter regulation, particularly regarding product safety, sustainability, and the environment. Currently, engineers use design software that was originally conceived in the 1960s. The new N8 facility, Bede, will allow researchers at the University of Manchester to develop the next generation of design tools, incorporating artificial intelligence, effectively re-engineering engineering.
|
|
Targetting of the NLRP3 inflammasomePrinciple Investigator: Richard BryceResearch Profile Project DescriptionBesides the beneficial events of inflammation, there are toxicological events arising from prolonged and uncontrolled inflammatory response that leads to tissue damage. Targeting of inflammasomes is a new approach for the treatment of inflammatory diseases. In this project, we have modelled the full-length human NLRP3 (NLR family pyrin domain containing 3) inflammasome. Molecular dynamics (MD) simulation will be used to study the structure and dynamical features of this NLRP3 monomer in water. Moreover, we will design novel inhibitors and MD simulation will be used to assess the interaction of the designed inhibitors with the target protein.Publications
|
|
Large-scale simulations of viral envelopesPrinciple Investigator: Jim WarwickerResearch Profile Project DescriptionDespite modern medicine advances, viruses remain a major health concern, both from a standpoint of occasional seasonal epidemics or major global pandemics. SARS-CoV-2, a member of the coronavirus family, has recently emerged as the cause of the most debilitating pandemic in the last hundred years. Flaviviruses, on the other hand, are mosquito-borne viruses (including dengue, Zika, yellow fever, and West Nile virus) which have been a major health concern in all regions within the Aedes mosquito range.Our group is focused on exploring the dynamics of coronavirus and flavivirus envelopes, which consist of protein, membrane, and sugar components, and which represent a protective coating for the virus, as well as being functionally important for viral entry into the host cell. We are using large-scale molecular dynamics simulations to investigate motions of coronavirus and flavivirus envelopes, including mapping of cryptic pockets to detect possible drug-binding hotspots. Publications
|
|
Atomistic simulation of liquid crystalsPrinciple Investigator: Andrew John MastersResearch Profile Project DescriptionA system is chiral if it cannot be superimposed upon its mirror image. One example is a left-handed helix, where the mirror image is a right-handed helix. Chiral molecules often form chiral structures, such as helices. It is rare, however, for achiral molecules to form chiral structures. Recent experiments have shown ( https://www.nature.com/articles/s41467-017-02626-6) that a series of achiral molecules can do just this, forming a nematic twist bend phase and a heliconical tilted smectic C (SmCTB) phase. We aim to perform molecular dynamics simulations to understand why these molecules form these phases and the predict the physical properties of these phases.
|
|
Classification of airborne biological particles in cities.Principle Investigator: David ToppingResearch Profile Project DescriptionPrimary biological aerosols (PBA) consist of bacteria, viruses, fungal and plant spores, fragments of plant or animal matter and pollen. Whilst their role in the environment, human health, and as a potential security threat is known, development of robust detection technologies remains a challenge. In this project,t we use multiple machine learning libraries to cluster images of individual particles into distinct types to better understand their sources, sinks, and potential impacts.Publications
|
|
Spatial-aware decision-making with ring attractors in reinforcement learning systemsPrinciple Investigator: Richard AllmendingerResearch Profile Project DescriptionThis project aims to integrate novel neural circuit-inspired models into reinforcement learning algorithms to enhance decision-making and uncertainty estimation in spatial scenarios. We target implementing these neural structures, which encode the spatial distribution of actions, both as standalone networks and as integrated components of deep learning models. The goal is to improve performance in challenging robotics and navigation benchmarks. Experiments will compare our approach to state-of-the-art baselines across various environments and action spaces. We expect this innovative approach to significantly speed up learning and lead to more robust decision-making, especially in partially observable settings.
|
|
Foundation Models for Nanoconfined WaterPrinciple Investigator: James McHughResearch Profile Project DescriptionThis project explores how water behaves when trapped in spaces just a few atoms wide – a state known as nanoconfinement. By using advanced machine learning techniques, we aim to build highly accurate models that predict the behaviour of water and dissolved ions under these extreme conditions. These predictions are based on quantum mechanical calculations, which are usually too expensive to carry out at scale. With the help of GPU-accelerated models, we can rapidly explore thousands of possible combinations of liquids, membranes, and ions. The results could help design better systems for water purification, selective ion extraction from electronic waste, and catalysis for green energy technologies. Ultimately, this work will enable faster, cheaper development of new materials and processes with real-world environmental impact.
|
|
Deep Constrained Clustering and Its ApplicationsPrinciple Investigator: Shaojie ZhangResearch Profile Project DescriptionOur project is dedicated to Deep Constrained Clustering (DCC) and Its Applications. There is ample evidence indicating that Constrained Clustering (CC), based on additional constraining priors, can significantly enhance clustering performance. Building upon this, the DCC paradigm, which utilizes Deep Neural Networks (DNNs), further augments the capability to discover patterns in large-scale, complex data. However, the extent to which existing DCC methods are influenced by the distribution of constraint information remains unclear. Our current work is endeavouring to uncover this impact and is attempting to circumvent potential negative effects through appropriate strategies.
|
|
Explainable Deep Reinforcement LearningPrinciple Investigator: Canh An Tien PhamResearch Profile Project DescriptionThe project aims to develop an explanation of the AI agent for users in decision-making problems (e.g. Atari Games). We are developing a model that has an attention map which can be visualized to demonstrate which locations the agent is focusing on to produce the actions. Furthermore, we also develop the action predictor model based on Generative AI to provide the estimation of agent actions that it could perform in the next few steps
|
|
Modelling opinion shifts of mental health groups via joint deep clustering and dynamic topic modellingPrinciple Investigator: Xian YangResearch Profile Project DescriptionThis research aims to analyze the impact of social events, like COVID-19 and vaccination programs, on the mental health of individuals. It proposes using AI and digital tracing data to understand the behavioural patterns of mental health groups. The study plans to address three key research questions: 1) Identifying consensus opinions among people with diverse mental health issues, 2) Analyzing the impact of events on opinion shifts, using dynamic topic modelling and NLP techniques, and 3) Mitigating bias in digital tracing data from various social media platforms. The research intends to validate its findings through surveys and interviews, offering potential insights for decision support systems in various domains.
|
|
Non-specific binding of adenosine triphosphate (ATP) with proteins and the impact on protein self assemblyPrinciple Investigator: Robin CurtisResearch Profile Project DescriptionAdenosine tri-phosphate (ATP) is the main energy metabolite consumed by cells to generate energy. Interestingly ATP is only required at micromolar concentrations in order to meet metabolic requirements, while cellular concentrations are one thousand times greater in the mM range. This discrepancy has led researchers to unravel a secondary role of ATP in diseased states of cells through preventing protein mis-assembly. However, the molecular basis for how ATP interacts with proteins and alters their assembly pathways remains poorly understood. We plan to tackle this problem using molecular simulations. Critical to the success of the project is to choose the correct forcefield which will be accomplished by benchmarking the simulations against experimental data derived from NMR spectroscopy. Once developed, the simulations will be extended to elucidate how ATP interacts with protein surfaces and the impact on protein-protein interactions. The project also has important implications for using ATP or ATP-mimic molecules as excipients for stabilizing biological medicines.
|
|
Machine learning for urban temperature estimates in the UK: a modelling testbedPrinciple Investigator: Zhonghua ZhengResearch Profile Project DescriptionUrban areas are particularly vulnerable to the effects of climate change and urbanisation, but they are also the foci for climate-resilient efforts, including adaptation, mitigation, and sustainability. Process-based simulations of urban climates are computationally expensive. Instead, data-driven approaches, such as machine learning (and Deep Learning), offer alternate approaches for estimating urban temperature. In this project, we will develop Deep Learning models for emulating urban climate (e.g. temperature) in the UK.
|
|
Distributed deep reinforcement learning with a group of agentsPrinciple Investigator: Xiao-Jun ZengResearch Profile Project DescriptionReinforcement learning has been an essential tool to tackle challenging problems, while it is non-trivial to design dense rewards properly for many challenging tasks. Alternatively, sparse reward settings in which the agent only receives valuable rewards for reaching some specific goals are easier to set up. For various tasks with different goals, they can be unified via formalisation as a goal-conditioned reinforcement learning (GCRL) problem. However, how to solve those GCRL problems efficiently under sparse reward settings remain challenging. In our work, we accelerate the training by decomposing the original task along existing trajectories into multiple sub-tasks that are easier to solve and we can assemble a sub-optimal solution from them for the agent to learn in the beginning. That makes learning faster.
|
|
Generative methods for subgrid turbulence closurePrinciple Investigator: Alex SkillenResearch Profile Project DescriptionDirect simulation of turbulent flows is typically too expensive, even with modern supercomputers. Coarser (and hence cheaper) meshes are employed in "Large Eddy Simulation" (LES). However, turbulent transport at scales finer than the mesh must be modelled in LES. Traditional turbulence models are empirical and tend to be inaccurate, particularly for flows where body forces promote turbulence anisotropy. In this project, we use generative methods in machine learning to generate sub-grid structures that approximately match the distribution of direct simulation. This is similar to upscaling images with “super-resolution” in the image processing field. In doing so, we hope to find an approximate inverse LES filter that can subsequently be used in improved LES modelling of plasma flows and liquid metal flows for the nuclear fusion industry.
|
|
Predicting the genomic risk of coronary artery disease through deep learningPrinciple Investigator: David TalaveraResearch Profile Project DescriptionCoronary artery disease (CAD) is the most common cause of death in the UK. Whilst the risk of developing CAD depends heavily upon lifestyle factors such as diet, exercise and cigarette smoking, genetics also play an important role. We aim to leverage deep learning approaches to improve upon the current methods used to predict people’s genetic risk of CAD, enabling targeted treatment interventions for those at the greatest risk of the disease.
|
|
Supervised learning of nonlinear superposition principle for hydrodynamic loading on structuresPrinciple Investigator: Ajay Bangalore HarishResearch Profile Project DescriptionStructures like offshore wind turbines, marine renewable energy structures, bridge piers, and floating vessels are routinely exposed to harsh environmental conditions. This is even more frequent during hazard events. These frequently drive the design. The physics and statistics of wave-structure interaction are complex and still not fully understood for strongly non-linear loads as experienced in the most severe conditions. In this work, we aim to demonstrate the possibility of a neural-network-based surrogate model capable of probabilistically predicting the drag coefficients on the structures. In this regard, a building array subjected to wave loading serves as a demonstration for the more complex city configuration to be considered in future studies.
|
|
Large language models in accounting and finance (A&F-LLMs)Principle Investigator: Eghbal RahimikiaResearch Profile Project DescriptionFinText (FinText.ai) is a collection of financial natural language processing (NLP) models that we developed in collaboration with the University of Oxford. We plan to develop FinText further with more advanced financial modules. All existing advanced NLP models are developed using only general textual information and, hence, are not precise enough for use in accounting and finance. We focus on incorporating high-quality textual accounting and financial resources to develop more reliable, state-of-the-art, robust, and transparent large language models (LLMs). We expect this new repository and related papers to become the primary foundation for future research in financial and accounting analysis, financial forecasting, ESG (Environmental, Social, and corporate Governance) monitoring, climate risk assessments, and business analytics.Publications
|
|
A Neural Network Based Surrogate to SWAN Nearshore Wave ModellingPrinciple Investigator: Lee MargettsResearch Profile Project DescriptionOffshore renewable energy generation relies on accurate and high-resolution wave model simulations for both device longevity and efficient energy production. Downscale wave modelling is an expensive task due to its high computational demand and the inaccessibility of software. The consequence is that access to offshore power development is limited to areas with a pre-established wave modelling community and access to high-performance computing. A machine learning approach can be adopted to simplify the modelling framework. Neural networks can be written to work in both GPU and CPU formats. With added simplicity and computational efficiency, machine learning-driven wave models can be better equipped to handle decades of data within a reasonable time frame. This flexibility renders increased accessibility within the industry by vastly reducing the computation time required by traditional wave models This project aims to demonstrate a surrogate neural network framework for the well-established SWAN (Simulating Waves Nearshore) model in the Outer Hebrides, UK.
|
|
Goal-conditioned Reinforcement LearningPrinciple Investigator: Ke ChenResearch Profile Project DescriptionReinforcement learning has been an essential tool to tackle challenging problems, while it is non-trivial to design dense rewards properly for many challenging tasks. Alternatively, sparse reward settings in which the agent only receives valuable rewards for reaching some specific goals are easier to set up. For various tasks with different goals, they can be unified via formalisation as a goal-conditioned reinforcement learning (GCRL) problem. However, how to solve those GCRL problems efficiently under sparse reward settings remains challenging. In our work, we accelerate the training by decomposing the original task along existing trajectories into multiple sub-tasks that are easier to solve, and we can assemble a sub-optimal solution from them for the agent to learn in the beginning. That makes learning faster.
|