|
Computer Vision - Advancing Automated Image UnderstandingPrinciple Investigator: Toby Breckonhttps://www.durham.ac.uk/staff/toby-breckon/ Project DescriptionThe work of my research team relates to computer vision and robotic sensing – the automatic interpretation of images by computers as an aspect of machine intelligence. Our work is enabled by both the fundamentals of image processing and recent advances in deep machine learning.Within this domain, we specialize in several industry-facing problem domains spanning X-ray image understanding, automotive vision (autonomous vehicles), visual surveillance, robotic sensing, and general topics in object detection, classification, and broader image understanding.
|
|
Deep Riverscapes: Advancing fluvial sciences with Deep learningPrinciple Investigator: Patrice Carbonneauhttps://www.durham.ac.uk/staff/patrice-carbonneau/ Project DescriptionLike all disciplines related to Earth Observation, Fluvial Remote Sensing (FRS) is currently benefiting from an explosion of data in the form of images from platforms such as drones and satellites. The Deep Riverscapes initiative aims to bring deep learning and artificial intelligence to bear on the range of theoretical and practical issues that currently face river science.
|
|
Electronic StructurePrinciple Investigator: Stewart Clarkhttps://www.durham.ac.uk/staff/s-j-clark/ Project DescriptionAb initio modelling of materials is employed in a wide range of research fields and plays a key role in the development of diverse technologies, and its importance in many research and technology areas is growing rapidly. Electronic structure methods are capable of modelling a very wide range of material systems, ranging from magnetism and superconductivity in complex oxides, where the advanced methods for treating the quantum many body problem must be applied, to very large systems such as the nature of crack propagation in alloys or studies of protein solvation, where we are now able to treat many thousands of atoms. A wide range of material properties can be calculated including optics and transport, nuclear magnetic resonance, phonon modes and magnetic excitations and these can be compared with experiment to allow a unique and unprecedented microscopic understanding of complex physical processes.
|
|
First Principles Simulation of Structure Property Relations in FerroicsPrinciple Investigator: Nicholas Bristowehttps://www.durham.ac.uk/staff/nicholas-bristowe/ Project DescriptionOften it is highly desirable to turn to a theoretical tool that does not rely on experimental parameterisation. Examples in the field of materials physics include designing new materials which have yet to be made, exploring the phase space of a material which has yet to be studied, or when attempting to unravel the underlying mechanisms behind unexpected observations. This project uses Density Functional Theory (DFT) to study the connection between the structure and properties of complex materials for potential next-generation devices.
|
|
Long document multitask learningPrinciple Investigator: Noura Al Moubayedhttps://www.durham.ac.uk/staff/noura-al-moubayed/ Project DescriptionThis project involves using machine learning to analyse long text documents such as books, movie scripts, and news articles. Current techniques for translation, automatic summarisation, and a range of other problems can only perform well on a few sentences or short paraphrases. We aim to develop techniques for allowing machines to understand documents of tens of thousands of words and reason usefully about them.This could be used to translate an entire book from French to English, detect news articles with bias, identify plagiarized sections in long pieces of work, summarize academic papers, and identify heroes and villains in movie scripts, alongside many other applications.
|
|
Emergence of sequence in a prebiotic mineral-assisted peptide growthPrinciple Investigator: Matteo Degiacomihttps://www.durham.ac.uk/staff/matteo-t-degiacomi/ Project DescriptionProteins are large biomolecules responsible for most life processes in any organism. They are constituted by simple molecular building blocks called amino acids, arranged into a chain. The specific amino acid order in the chain determines a unique protein’s three-dimensional shape, which in turn determines its biological function. Interestingly, the biological machinery dedicated to producing proteins in a living organism involves itself, proteins. The question is thus, how did the first proteins come to be? In our previous work (Erastova et al., Nature Communications, 2017), we have demonstrated that, in Early Earth, mineral surfaces may have promoted the spontaneous formation of amino acid chains. In this project, we will investigate the potential of mineral surfaces to promote the formation of specific amino acid sequences.
|
|
Analysis of the Credibility Revolution in EconomicsPrinciple Investigator: Nejat Anbarcihttps://www.durham.ac.uk/staff/business-staff/nejat-anbarci/ Project DescriptionThis project examines the evolution of topics in economics in particular the movement of economics away from theory to empirical work. This project will utilize state-of-the-art machine learning methods in order to classify papers in economics by both their field and the methodology used in order to determine where this change has been greatest. Other parts that this project will examine is the relationship between changes in economics and educational institutions, as well as the relationship between categories of economic journals. Furthermore, this project will also examine factors that are related to the production of certain types of research in economics.
|
|
Recognizing Human-object interactions in videosPrinciple Investigator: Frederick Lihttps://www.durham.ac.uk/staff/frederick-li/ Project DescriptionA human activity typically involves a series of various human-object interactions. Understanding human activity, it usually involves the identification of individual human actions and the objects the which humans are interacting with. Such tasks are typically challenging due to the complication of spatial relationships among human and objects in a scene, and the possible changes or inconsistency of temporal relationships between humans and objects.Our project aims to investigate deep learning methods to recognize human-object interactions (HOIs) from a video sequence. Our approach is to learn both spatial and temporal features from video sequences, making use of such features to infer a latent space that can well discriminate different classes of HOIs. Applications of our findings include activity prediction, motion planning, video surveillance, etc.
|
|
Computer simulations of soft matter systemsPrinciple Investigator: Mark Wilsonhttps://www.durham.ac.uk/staff/mark-wilson/ Project DescriptionMolecular simulation provides a microscopic "picture" of how molecules interact and self-organise. It can be used to understand how small changes in molecular structure can lead to important changes in how molecules interact with each other. This, in turn, influences how molecules self-organise to form complex systems (for example, proteins and membranes), and controls complex phase behaviour (such as controlling the formation of liquid crystal phases). The advanced molecular simulations possible today, using techniques such as "molecular dynamics" and "dissipative particle dynamics", can be used to help researchers understand these complex behaviours of molecular systems, and are helping researchers plan new experiments and design new molecules to have specific functions.
|
|
Leverage Transfer Learning for Efficient Generative Models for Natural LanguagePrinciple Investigator: Alexandra I Cristeahttps://www.durham.ac.uk/staff/alexandra-i-cristea/ Project DescriptionUnderstanding natural language in-depth is one of the important and timely goals of Artificial Intelligence. Generative models are powerful techniques to model information in a probabilistic manner. They help to represent information that may not have been available to the model when training, so basically extrapolating to unseen data. Here, we aim to apply generative modelling with modern state-of-the-art architecture to better understand human language, by using the Bede supercomputer.
|
|
Large-scale generative modellingPrinciple Investigator: Chris G. Willcockshttps://www.durham.ac.uk/staff/christopher-g-willcocks/ Project DescriptionThis project investigates training a large generative model on protein data (10,000+ proteins) using recent variants of linear Transformers. In particular, we estimate the next atom in a sequence conditional on previous atoms and other observable data. At inference time, we sample from a continuous space of probable sequences, generalising to new biologically plausible configurations unseen during training. Additionally, we enforce known physical constraints at the intermediate protein positions, leading to further improved generalisation performance and training robustness. This project is a collaboration between Durham University Department of Computer Science, the Department of Physics, and the Department of Chemistry at the University of Uppsala.
|
|
AI-Powered Laboratory Assistant for ChromatographyPrinciple Investigator: Rui Pedro Carvalhohttps://www.durham.ac.uk/staff/rui-carvalho/ Project DescriptionThis project aims to develop a locally deployable AI system for real-time chromatography analysis at the UK Centre for Process Innovation (CPI). We will train small, efficient language models to function as automated lab assistants capable of analysing chromatography data, identifying experimental issues, and performing preliminary data interpretation during experiments, bridging the critical time gap between experiment execution and expert analysis.
|
|
Discovering Emergent Symmetries via Machine LearningPrinciple Investigator: Enrico Andriolohttps://www.durham.ac.uk/staff/enrico-andriolo/ Project DescriptionThis project applies cutting-edge machine learning techniques—specifically Graph Neural Networks (GNNs)—to a longstanding challenge in theoretical physics: uncovering hidden symmetries in lattice field theories. These symmetries, while essential to understanding the underlying laws of nature, often do not appear explicitly in the equations that describe these systems. By training GNNs on large-scale simulations of these lattice models, we aim to detect and characterize such symmetries automatically, potentially revealing new physical insights. This work sits at the intersection of artificial intelligence and fundamental physics and may pave the way for machine-assisted discovery in complex scientific domains. Access to high-performance GPUs is critical to handle the computational demands of training these models on large and intricate datasets.
|
|
Braced excavations: what about the corners?Principle Investigator: Mao Ouyanghttps://durham.ac.uk/staff/mao-ouyang/ Project DescriptionAn EPSRC-funded project looking at large braced excavations, with Professor Charles Augarde as PI, Will Coombs and Alexandros Petalas as Durham CoIs and Jonathan Knappett and Mike Brown at Dundee University. Durham is doing numerical and Dundee experimental. Started in July 2023 for 4 years. Project partners are Laing O’Rourke, AKT II, Oasys, Cementation Skanska & Southampton University. Durham PDRA is Mao Ouyang and Dundee PDRA is Ahmed Alagha.
|
|
Atomistic simulations of rhamnolipids in lipid bilayersPrinciple Investigator: Mark Millerhttps://www.durham.ac.uk/staff/m-a-miller/ Project DescriptionSurfactants are a type of molecule that can interact both with water and with materials that are normally insoluble, like oils. Many consumer products, such as washing-up liquid, rely on surfactants and they are manufactured in great quantities. Biosurfactants also have this "amphiphilic" property of interacting with water and oil but are naturally occurring and are synthesised by certain bacteria. They have the advantages of being possible to produce from renewable sources and showing low toxicity. In this project, we are investigating a class of biosurfactants called rhamnolipids, which contain one or two rhamnose groups (a type of deoxy sugar). We are developing accurate computational models to predict and understand the physical properties of rhamnolipids with a view to safety screening and possible application in environmentally friendly products.
|
|
Machine learning classical density functional theoryPrinciple Investigator: Stephen Coxhttps://www.durham.ac.uk/staff/stephen-j-cox/ Project DescriptionMolecular dynamics and Monte Carlo simulations are powerful techniques to understand the behaviour of liquids. However, they can be computationally expensive, and much of the time, the exact details of what individual atoms or molecules do are unnecessary to understand the problem at hand. Approaches rooted in classical density functional theory (cDFT), on the other hand, provide the average equilibrium density and free energy simply by minimising a free energy functional. However, the free energy functional is in general not known.In this project, we will use machine learning techniques to establish the free energy functional for various liquids, including water, molten salts and mixtures. This will pave the way to understanding liquids and liquid-solid interfaces accurately and at much reduced computational cost.
|
|
Methods for Advanced Statistical Machine LearningPrinciple Investigator: Viet Cuong Nguyenhttps://www.durham.ac.uk/staff/viet-c-nguyen/ Project DescriptionThis project aims to develop new mathematical theory and methods to train large-scale machine learning and artificial intelligence models. We will consider different statistical learning frameworks such as continual learning, transfer learning, and Bayesian inference, and develop suitable training methods that can achieve state-of-the-art performance on computer vision and natural language data.
|
|
A machine learning-aided investigation of potential relationships between feelings of belonging or exclusion and social attitudesPrinciple Investigator: Patrick Kotzurhttps://www.durham.ac.uk/staff/patrick-f-kotzur/ Project DescriptionInterpersonal relationships afford a range of sociopsychological benefits, from a sense of belonging to support during difficult times. Social exclusion has, however, been reported to be increasing in the United Kingdom, alongside levels of partisan and intergroup polarisation. This project aims to utilise topic modelling and machine learning classifier algorithms to identify potential links between feelings of belonging or exclusion and social attitudes in a UK context.
|
|
Empowering Power Electronics: A Language Model Approach to Design OptimisationPrinciple Investigator: Mahmoud Shahbazihttps://www.durham.ac.uk/staff/mahmoud-shahbazi/ Project DescriptionLarge language Models (LLMs) like ChatGPT and Google Gemini have recently demonstrated exceptional capabilities. This innovative project aims to revolutionize power electronics design by leveraging cutting-edge LLMs. By training a specialized LLM, we aim to empower engineers and researchers in the field of power electronics to make informed decisions swiftly and accurately. The developed model will not only provide guidance on selecting optimal converter topologies, semiconductor devices, and control strategies but also offer insights into emerging technologies and design trends. Ultimately, this project seeks to enhance efficiency, reliability, and performance in power electronic systems, driving advancements in renewable energy, electric vehicles, and beyond.
|
|
Topological Data Analysis of High Volume Data SetsPrinciple Investigator: Fernando Galaz-Garciahttps://www.durham.ac.uk/staff/fernando-galaz-garcia/ Project DescriptionThis project employs topological data analysis (TDA) to examine the shape of point clouds in Euclidean space, focusing on phenomena of relevance in political science. Point clouds represent data points in a high-dimensional space, and TDA provides a mathematical framework to extract meaningful insights from their shape and structure. The project aims to uncover hidden topological patterns and structures within the point clouds that correspond to political phenomena of interest by applying TDA techniques, such as persistent homology. This analysis can provide valuable insights into the relationships, clustering, and dynamics of political events, such as voting patterns.
|
|
Prediction of protein structure using N8 HPC and alphafoldPrinciple Investigator: Wenbin Weihttps://www.durham.ac.uk/staff/wenbin-wei2/ Project DescriptionThis project will install and apply alphafold (https://github.com/deepmind/alphafold) to support researches that require the prediction of protein structure in the Bioscience Department of Durham University.
|
|
Using deep learning approaches to map Greenland Ice Sheet fracturePrinciple Investigator: Thomas Chudleyhttps://www.durham.ac.uk/staff/thomas-r-chudley/ Project Description"The Greenland Ice Sheet is the largest contributor to sea level rise from glaciers and ice sheets. Meltwater, which flows through the ice sheet to the oceans, can accelerate ice loss by lubricating the ice bed and driving melt at the ice front. However, these impacts are controlled by how water enters the ice sheet, and to date no one has mapped where and how crevasses are transferring water the bed. I will use Bede to train a deep learning classifier to map crevasse hydrology across the ice sheet from satellite data. I will implement these discoveries into a coupled model of ice hydrology and dynamics, exploring of crevasse hydrology on Greenland’s future for the first time.For more information, see https://www.leverhulme.ac.uk/early-career-fellowships/how-do-crevasses-transfer-water-bed-greenland-ice-sheet
|
|
A molecular dynamics investigation into the solid-liquid interfaces under A/C fieldPrinciple Investigator: Mehdi Tavakolhttps://www.durham.ac.uk/staff/mehdi-tavakol/ Project DescriptionInvestigating the organization and dynamics of ions at solid-liquid interfaces is vital for advancements in science and technology. While previous studies have focused on the spatial distribution and dynamics of ions, the lateral distribution near these interfaces has been largely overlooked. This knowledge gap hinders our comprehensive understanding and restricts potential applications. To bridge this gap, our research project employs molecular dynamics simulations to explore the lateral distribution of ions at solid-liquid interfaces. By capturing this additional information, we aim to unravel the complex behavior and mechanisms underlying interfacial phenomena. The insights gained from this research have the potential to contribute significantly to fundamental knowledge and open doors to innovative technological developments in fields ranging from materials science to energy storage and beyond.
|
|
Mechanistic insight into the occurrence characteristics of nano-confined fluid in clay mineral and organic matter pores through molecular dynamics simulationPrinciple Investigator: Prof. H. Chris Greenwellhttps://www.durham.ac.uk/staff/chris-greenwell/ Project DescriptionClay minerals are ubiquitous in the natural environment. They are present in the soil system, in sediments, in the sub-surface, and even within the air. Clay minerals have unusually high surface area and thus absorbent properties, taking up many different types of nutrients, pollutants, and naturally occurring fluids. The type of fluid already at the surface will determine the other, incoming, fluids relative affinity for the surface (this is known as wettability alteration). Understanding the process by which one fluid replaces another is key, and this requires molecular-level understanding only available by molecular simulation.
|
|
Understanding membrane breakdown at anti-microbial polymer surfacesPrinciple Investigator: Mark Wilsonwww.dur.ac.uk/mark.wilson Project DescriptionRecent experimental work has indicated that carefully treated surfaces can act as anti-microbial surfaces, with many potential benefits in areas such as food preparation and health care. This project is about understanding, at a molecular level, the mechanisms involved with these surfaces and, in particular, how they can lead to cell membrane breakdown. The models to be used are coarse-grained molecular models that capture the molecular interactions between representative model membranes and treated polymer surfaces, but are sufficiently computationally tractable to allow the long time scales required to allow membrane damage to be simulated.
|
|
Potential Impact of Open Banking on the Development of Islamic Fintechs: Portfolio Optimization with Machine Learning as An Islamic FinTech ApplicationPrinciple Investigator: Gokmen Kilichttps://www.artvin.edu.tr/gokmen-kilic Project DescriptionThe main purpose of this research is to bring together the investment instruments of six participant banks operating in Turkey under a single web platform that enables distributed deposits and investment instruments to be managed through this platform and allows choosing the best portfolio for the customer. Technically, this platform can benefit from all advantages open banking API integration system framework as possible the banks’ API markets allow us. This means the customer reaches out to any investment instruments of six participant banks and can make investments through this platform. This research approach to portfolio selection in two financial models to figure out which portfolio is best for the customer. The first financial model is Markowitz's model, known as the Modern Portfolio Theory for portfolio selection, according to predetermined criteria. The second financial model is machine learning financial algorithms that include various techniques to be able to maximize return and minimize risk by using Python financial libraries and other coding techniques. Thus, comparing the performances of these two financial models shows us which model is effective in terms of risk and return of the portfolio.
|
|
JUNEPrinciple Investigator: Frank Krausshttps://www.ippp.dur.ac.uk/profile/krauss/ Project Description"JUNE is an epidemiological simulation program that models and projects the spread of an infectious disease such as COVID-19 through large-scale virtual populations. Constructed during the ongoing COVID pandemic, it focuses on human-to-human transition pathways. In the original version a digital twin of the about 55 million individuals of the English population was constructed, based on geographically and demographically highly granular ONS data, which provided insights at the level of on average 250 residents, and allows us to map out their daily lives and, most importantly, their contacts. Subsequently, JUNE has also been used to project the progress of the pandemic in other populations, among them the residents of Cox's Bazaar, one of the world's largest refugee camps with about 800,000 inhabitants. In this project we aim to port JUNE to a GPU cluster, which will necessitate parallelization of the code, reflecting the size of the population."
|
|
Percolation of Active Brownian ParticlesPrinciple Investigator: Mark Millerhttps://www.durham.ac.uk/staff/m-a-miller/ Project DescriptionAny chemical solution or suspension of nanoparticles contains transient clusters. The average cluster size increases with concentration and may also be affected by other conditions, such as temperature. At the so-called percolation threshold, the clusters suddenly grow from microscopic to spanning the entire sample. This transition can be accompanied by a dramatic change in material properties. For example, percolation of electrically conducting particles transforms an insulator into a conductor. We are investigating what happens when the clustering particles are “active” – i.e., they propel themselves by consuming energy. Active matter is a major topic in modern research, covering everything from self-propelling colloids to bacterial cells, but percolation in active matter has not been addressed so far. Existing theories for traditional “passive” materials cannot be applied. Computationally demanding simulations are needed to investigate this uncharted territory.
|
|
Synthetic data in the social sciences using generative adversarial networksPrinciple Investigator: Ajeeth Kanagarjanhttps://markwilson.webspace.durham.ac.uk/people/ Project DescriptionSolubility of small molecules in polymeric materials is important in a variety of sectors, including packaging, adhesives, medical, cosmetics, and food. Obtaining a very high solubility for a certain chemical is important for production. Polyvinyl acetate is one of the most important polymers in these industries. We are developing computational models such as atomistic and coarse-grained models for these polymers and developing methods for predicting the solubilities of other molecules in this polymer using classical simulations.
|
|
Lattice Boltzmann simulations of turbulent flows interfacing with porous mediaPrinciple Investigator: Halim Kusumaatmajahttps://www.durham.ac.uk/staff/halim-kusumaatmaja/ Project DescriptionTurbulent mixing and transport are relevant for wide-ranging natural and engineering processes, from improving the performance of heat exchangers and oil recovery processes to mixing within river sediments. Our focus here is on the turbulent boundary layer that is formed at the interface of a free flow and porous medium, where infiltration of turbulence from free flow into the porous layer can lead to highly complex mass, momentum, and heat transport. However, a fundamental understanding of mixing and transport processes at the boundary layers is still lacking. To make progress, it is important to better understand the complexity of the flow dynamics that occur at several length and time scales, and this is the insight we will generate using a powerful computational fluid dynamics (lattice Boltzmann) simulation method.
|
|
Transformers for Ancient Greek and LatinPrinciple Investigator: Prof. Peter Heslinhttps://www.durham.ac.uk/staff/p-j-heslin/ Project DescriptionThe field of Natural Language Processing was revolutionised by the introduction in 2017 of the Transformer architecture for large language models, such as BERT. This project uses that same architecture to train models for Ancient Greek and Latin. This can provide a way of testing our existing digital texts for anomalies. Where the model fails to predict the existing text, we can look to see if it is a genuinely exceptional linguistic usage or if it represents an error: either in the digitisation, or in the printed edition the digital version was based upon, or even in the transmission of the text by means of generations of handwritten manuscripts over the millennia. Another research question is how best to train a large language model with very limited amounts of linguistic data. This aspect of the project may help to shed light on the problems faced by all low-resource languages, including languages spoken in the developing world.
|