The cluster welcomes interest from anyone who is looking pursuing a PhD in our areas of interest, which broadly encompass statistics, probability, and mathematical biology. We recommend that you start with an informal approach to potential supervisors who work in areas of interest to you; they will be able to discuss potential projects and give you advice about applications. You can then apply here.
Our cluster is advertising four PhD projects under the University's Doctoral Landscape Award programme, which will be fully funded for successful candidates. Applications are now open and the deadline is mid-January.
Full details on these projects and funding can be found here.
For fifty years, computer chips improved by packing more components onto one piece of silicon. That route is closing. Systems are now built by combining several smaller pieces of silicon, each made in the technology that suits it best, inside a single package. This approach powers data centres, electric vehicles, medical scanners and phones, and it is a stated priority of the United Kingdom's semiconductor strategy.
It also creates a hard design problem. The decisions that remain manual sit above the steps software already handles: how to split a system into pieces, how to deliver power across them, how to lay them out so heat escapes, and how to judge whether an assembled system will work. Each choice trades speed, power, temperature, cost and reliability against the others. To judge a candidate design, engineers run simulations that take hours or days. With a few dozen simulations affordable and an astronomically large number of possible designs, most of the space is never examined.
Experienced engineers cope by knowing where to look. Consider a small company designing its first power management system for an electric vehicle. A veteran who has built twenty similar systems carries expectations about what tends to work, and spends the simulation budget accordingly. A newcomer has none, and this is a main barrier to entry in advanced electronics.
This project will develop artificial intelligence that builds those expectations from data rather than from a career. The idea is to learn from a large collection of past designs and simulations of related systems, and to distil that experience into an informed starting point for a system nobody has seen before. Crucially, the method must be honest about how much it knows. A confident but wrong expectation is worse than none at all, because it sends the search in the wrong direction and hides the failure until manufacture.
The student will work on three linked questions. First, how to compress experience from many past design tasks into a starting point that reports realistic uncertainty on a new task. Second, how to decide which design to simulate next and how accurately, in a way that stays safe when the learned expectations mislead. Third, how to notice from a handful of results that the expectations are wrong, and correct course.
Industrial data on these decisions is largely unavailable, so the project uses open reference designs and simulation data generated in-house, including open design patterns being produced by the national centre for multi die system design at Sheffield. A method is accepted only if it holds up on test problems chosen to be unlike those it learned from, since that is what a real new design presents. The final year connects the components to a working design procedure and measures how much simulation effort they save compared with current practice.
The project sits between statistics, machine learning and electronic engineering, and the student will be trained in all three. The supervisory team combines statistical expertise in uncertainty and decision making with an active research programme in machine learning for chip design, and the student can engage with the national centre for multi-die system design at Sheffield, which brings contact with the companies facing these problems.
The underlying question, how to reuse experience when data is scarce and experiments are expensive, recurs across engineering, drug discovery and climate modelling, so the skills gained are portable.
No prior knowledge of chip design is required. The project suits a mathematically minded graduate in mathematics, statistics, computer science, physics or engineering.
Infectious diseases of trees such as ash dieback, Dutch elm disease and sudden larch death are devastating UK woodlands—costing the economy billions, threatening biodiversity, and undermining climate change mitigation efforts. It is therefore crucial that we develop scientifically-evidenced management strategies to better control disease spread.
This interdisciplinary PhD project sits at the intersection of mathematical and statistical modelling and forest ecology. Whether you are a mathematician looking to apply your theoretical skills to real-world ecological problems, or a quantitative biologist eager to develop advanced modelling skills, this project offers you the opportunity to use modelling approaches to shape forest management. You will develop and analyse predictive epidemiological models to understand disease spread, with a focus on Phytophthora ramorum, the causal agent of sudden larch death. You will then use these models to advise on control strategies that might protect woodlands against disease. The project is in collaboration with Defra (Department for Environment, Food & Rural Affairs) - including a 3-6 month placement - providing a pathway for direct application of your findings.
There is a rich history of mathematical models of disease, including their prominent role in understanding the COVID-19 pandemic. An important element of plant diseases is that host individuals are usually fixed in their spatial location, with neighbour-to-neighbour spread dominating. Appropriate spatially-explicit mathematical models are challenging to construct, but recent work in the field has successfully fitted spatial mathematical models to real data on disease spread. This opens the door to using these models to test different management and mitigation strategies and therefore advise forest managers on best practice.
The initial focus of the PhD project will be on developing mathematical and computational models, in the form of ordinary differential equations and individual-based stochastic simulations. These models will be tailored to reflect the biology of the spread of sudden larch death, particularly its localised spread, working closely with the co-supervisor from Defra to ensure the models are appropriate. You will then use advanced statistical approaches to fit the models to real data of disease spread and quantifying uncertainty in the predictions. Finally you will use your tailored, parameterised models to test different control strategies. There is also the opportunity to develop a user-friendly web interface for end-users to explore and interact with your modelling.
In many areas of science, there are multiple useful ways to model a given complex system, which may include simple descriptive models or more elaborate mechanistic or process-based models. A collection of different models, used in combination to give predictions within a formal mathematical and statistical framework, is known as a multi-model ensemble.
Recent research at the University of Sheffield has led to flexible methodology for creating and fitting a multi-model ensemble that allows for the different strengths and weaknesses of models for different types of prediction, rather than simply averaging or weighting the models. It involves constructing a statistical model of the discrepancies between models and between each model and the true, unknown underlying process. The methodology accommodates models with widely differing scope or resolution, or different representations of the system. For example, in ecology, it has been applied to combine models defined at species level with those in which species are merged according to their function in the ecosystem, and to combine multi-species models with one or more single-species models.
The aim of this project is to develop, implement and explore mathematical and statistical extensions to this methodology. In particular, there is a widespread scientific need for a statistical framework that can incorporate spatially explicit models within an ensemble, for example in regional climate models and species distribution models. At present, a spatially explicit model can be included only by aggregating over space within the model, losing information in the process. We aim to develop a framework allowing the spatial structure to be used in the ensemble itself, which would make better use of available information and allow spatially-resolved comparisons and predictions. We will apply these methods to questions in marine ecology and potentially other environmental applications.
The current ensemble approach represents the state of the art for predicting marine ecosystems, for both scientific and management purposes. It is widely used by the UK Centre for Environment, Fisheries and Aquaculture Science (Cefas) which is responsible for official advice on fisheries to the UK government. Cefas will be a formal partner in this PhD project; the supervisor there was involved in the original methodological research. The extensions will address specific environmental scenarios and management options of direct interest to Cefas, leading to better understanding and management of marine ecosystems and fisheries. The project is a unique opportunity to develop novel statistical methodology, with a wide range of potential uses as well as motivation from an important practical application, and with input both on ecology and on modelling from scientists responsible for official advice on fisheries to the UK government.
This 3.5-year studentship is part of the School of Mathematical and Physical Sciences at the University of Sheffield, within the Probability Theory research group. The project focusses on network archaeology in random trees, particularly preferential attachment trees with vertex removal. Network archaeology studies the problem of reconstructing the temporal history of a network when only its final state is observed. Understanding network evolution provides insight into its present structure and has applications in network science, biology (e.g. protein–protein interaction networks), and computer science. From a mathematical perspective, network archaeology presents interesting combinatorial, graph theoretical, and probabilistic problems, and hence has recently seen a wealth of activity in the (combinatorial) probability community, mostly in the setting where the underlying network dynamics generate trees.
Most mathematical work on network archaeology for random trees focuses on root finding, using only the observed tree structure. Here, given a tree T and accuracy epsilon, the goal is to output a (minimal) subset S of vertices in T such that the probability that the root is within the set S is at least 1-epsilon. The size K=K(epsilon) of the set S is known as the budget. Optimality of the order of the budget K in terms of epsilon is a further goal. Interestingly, for many different tree models there exist algorithms that require a finite budget K, irrespective of the size of the tree.
The studentship will generate impact by extending current state-of-the-art research regarding root finding to the more general setting of preferential attachment trees with vertex removal, as well as by tackling new problems within network archaeology on random trees. Students are valued and active members of the School and will be part of a wider network of PhD students. You will have the chance to influence and lead school activity, seminars and events, and meeting leaders in the field. The university offers, in addition to supervision, rigorous research and scholarly training throughout your PhD, as well as funding to travel to and attend summer schools, workshops, and conferences. The probability group at the university of Sheffield has close contacts with other research groups within the UK, for example via the `Probability in the NorthEast’ consortium. The university also hosts the Applied Probability Trust, thus leading to an increased activity in terms of event organisation.
Beyond academia, Sheffield is a medium-sized city, notably the greenest city in the UK, offering reasonably affordable housing, a vibrant city centre with a large independent scene, and easy access to nature in the Peak District.
Below are some other current projects without specific funding on offer within the research cluster.
Diseases and pests of trees lead to large economic costs to the forestry industry, limit natural spaces for recreation and impact climate change by reducing trees’ ability to capture carbon. A key element of infection specific to plants is spatial structure: since plants are usually fixed in location, infection tends to be localised. In this project you would develop advanced, spatially-explicit mathematical models using nonlinear ordinary differential equations and computational individual-based models. You will fit your models to real data of disease spread and then test different management and mitigation strategies.
Antagonistic coevolution between parasites and the hosts they infect is a major driver of disease dynamics. Mathematical models have been used for some time to understand these co-evolutionary dynamics, however, they tend to assume constant environments and no external interactions with other species. In this project you will develop coevolutionary models of hosts and parasites within the adaptive dynamics framework, primarily using nonlinear ordinary differential equations and some stochastic numerical models. You will incorporate complex environments either through fluctuating parameters (for example representing seasonality) or external interactions (such as competiiton with other species).
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Network archaeology studies the problem of reconstructing the temporal history of a network when only its final state is observed. Most mathematical work on network archaeology for random trees focuses on root finding. Here, given a tree T and accuracy \epsilon, the goal is to output a (minimal) subset S of vertices in T such that the probability that the root is within the set S is at least 1-\epsilon. The project will investigate root finding and other problems in preferential attachment trees with vertex removal. This random tree model generalises popular models such as affine preferential attachment and uniform attachment by including vertex removal. In the PAVER model, starting from a single root vertex, a random tree is constructed recursively by, at every step, either connecting a new vertex to a randomly selected vertex in the tree or by removing the selected vertex from the tree. The project will (1) extend known results for root finding in preferential attachment and uniform attachment models to the more general setting of PAVER models, and (2) develop and investigate algorithms to find the oldest unremoved vertex in the tree, a generalised root finding problem that is novel and unique to PAVER models (compared to models without vertex removal).
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
Random graph models are probabilistic models for the development of networks. They are often inspired by networks found in areas such as computing, social networks and biology, but the emphasis is on studying mathematical properties of the graphs. One family of random graph models which has been studied extensively for the last 25 years or so is the preferential attachment model and its variations. Some of these variations feature an interesting phenomenon often referred to as condensation, where a small proportion of vertices become unusually important in the network in that a large proportion of edges connect to them. The aim of this project will be to study this phenomenon further and its occurrence in new graph models.
The location of individual animals and species emerge from a complex combination of their interactions with one another and with the underlying environment. This combination makes it difficult to predict how species distributions may change as the landscape is altered (e.g. by human infrastructure or the effects of climate change). Mathematical models can help by providing a rigorous framework for understanding how the mechanisms of between-species interactions and animal-landscape interactions lead to the emergent patterns of location distributions observed in nature. This project seeks to further that understanding by analysing pattern formation in a class of partial differential equation (PDE) systems in heterogeneous environments. The technical difficulty arises from the fact that most PDE pattern formation studies assume a homogeneous environment, but this study will build on recent techniques of the supervisor for dealing with heterogeneous environments. If you are interested in this project, please contact the supervisor on the above link for more information.
This project studies a recently-proposed class of PDE systems for studying the spatial arrangement of interacting populations of animals (and, in some situations, cells). Numerical investigations have revealed rich pattern formation properties in these systems, whereby continuous changes in parameters can cause qualitative changes in spatial patterns (bifurcations), which in some circumstances are discontinuous (tipping points). This project will aim to understand these bifurcations and tipping points analytically. We will use traditional tools, such as Crandall-Rabinowitz bifurcation theory and weakly non-linear analysis, to analyse these phenomena, but also explore the extent to which we can ease the complicated calculations that arise, through careful use of large language models. Such usage is quite novel and has only been made potentially possible through the very recent rapid developments in LLM technology. Please contact the supervisors on the above links for more information.
Index trackers (ETFs, index funds, index mutual funds) aim to replicate the return of a reference equity index while keeping implementation frictions---notably turnover and trading costs---under control. Index tracking is a passive investment approach that frequently beats their actively managed counterparts over the long term. In practice, full index-replication can be costly or operationally infeasible; consequently, pure tracking with a sparse basket is common for segregated mandates, often subject to constraints such as a full-investment budget and, in some mandates, long-only weights. A large body of work formulates index tracking as an optimisation problem with convex losses and penalties, or as mixed-integer programs that directly control support size and transactions. These approaches typically deliver a single allocation weight vector and tune regularisation hyperparameters by heuristics or cross-validation. Such tuning requires manual interventions and is often subject to uncertainty.
This project develops novel Bayesian methodology with the aim to return a posterior over weights, tuning parameters, and the residual variance that drives tracking error. The aim is decision-grade uncertainty quantification (UQ) and principled data-driven parameter selection. Hence, the project will develop automatic portfolio rebalancing driven by UQ analysis, based on the posterior distribution of the allocation weights.
We will model the tracking residual returns—index minus portfolio—as a noisy outcome with unknown variance, regressing the index returns on the returns of the individual assets. Portfolio weights live in a constrained space reflecting business rules: non-negativity or signed positions, leverage or budget limits, optional sector exposures, and explicit cardinality. Sparsity comes from imposing hierarchical priors that prefer few active names: either spike-and-slab families (selection indicators plus a continuous “slab” for active weights) or continuous shrinkage alternatives such as horseshoe or Dirichlet-Laplace. Tuning parameters are not fixed numbers in this approach: global and local shrinkage scales, inclusion probabilities, tracking-error and turnover multipliers, and the residual variance all carry their own priors. The project is expected to result in a coherent hierarchy that links portfolio structure, transaction costs, prior beliefs, and a data-driven narrative for model selection.
Inference combines two ingredients. First, state-of-the-art MCMC samplers, which are designed to handle non-smooth penalties and hard constraints. These samplers perform gradient-informed random walks while enforcing feasibility at every step. These take advantage of recent advances and contribute to convex nonsmooth optimisation and numerical analysis of stochastic differential equations. Second, Gibbs or Metropolis updates for the hyperparameters: residual variance, global shrinkage, local scales, and selection indicators.
Alongside expert-set priors, the project proposes to also use machine learning techniques to learn parts of the prior and penalty structure from data in an interpretable way. Examples include mapping liquidity and volatility features to a prior preference for sparsity, learning sector-level shrinkage from historical co-movement, or learning a turnover multiplier that depends on spread and market depth. The plan is to start with simple parametric maps and progress to neural networks with monotonicity or convexity constraints so that learned priors remain stable and defensible. Expert ranges can be encoded as hyperpriors, allowing the data to refine but not overrule domain knowledge.
Please email the supervisor for more information
Please email the supervisor for more information
Please email the supervisor for more information
The cost and duration of traditional clinical trials imply a high societal and economic cost, hampering the release of new treatments to public. One possible avenue to improve on this situation is the use of mathematical models to produce synthetic data, which can then be shared with contemporary clinical trials to reduce the number of patients and the duration of the trial. The approach can take advantage of well established mathematical models of diseases, or use current information to generate statistical forecasting models for less well understood systems. A Bayesian approach allows combining disparate sources of information, along with coherent uncertainty propagation, fundamental to feed into validation, verification and uncertainty qualification needed for clinical approval of these novel medical devices.
Panel (longitudinal) data enables learning the dynamics and relations of (groups of) units, strengthening the inference on both cross-sectional and dynamic parameters. The dominant approach to modelling unbounded continuous quantities of interest is to assume symmetric error terms (eg Gaussian or Student), but this assumption is often not met in practice. The idea is to explore, from a Bayesian perspective, the use of skew error distributions from well-known skewing mechanisms (eg Azzalini, 1985; Ley 2010), while allowing the parameter(s) controlling the skewness to vary over time. This implies the design of a stochastic path to control the process in time and exploring conditions for convergence.
The Brownian web is a dense system of coalescing Brownian motions in 1+1 dimensional space time. It is known to be the universal scaling limit of a wide variety of models, ranging from population genetics and interface growth to theoretically important processes such as orientated percolation and directed spanning forests. The project will develop scaling limit theory for dense interacting systems of particles in cases where the limiting particle motions are not simply Brownian motions e.g. they might feature spatial inhomogeneity or heavy tailed behaviour.
Please email the supervisor for more information
Large random trees have well-known discrete and continuous limits which have been studied since the 90's. The aim of this project is to investigate how much changes if we force the trees to be in unusual regimes.
Life cycle assessment underpins environmental decision-making across industry and policy, but its practical use is limited by incomplete process data, the manual effort required to build and validate models, and the difficulty of adapting existing assessments to new materials or products. This project develops AI methods along three complementary lines. The first uses large language model based reasoning agents to automate parts of the assessment workflow itself, including reading technical documentation, navigating supply chains, and selecting appropriate modelling assumptions. The second develops probabilistic models, including neural processes and knowledge graphs, to complete missing inventory data with calibrated uncertainty when primary data is unavailable. The third uses active learning to prioritise which data gaps most affect the reliability of an assessment. The methods will be validated on industrial case studies with partners including AMRC and international collaborators, and the underlying approach extends naturally to broader circular economy questions such as material flow tracking and end of life material reuse.
Engineering design problems, from analog circuit sizing to physical simulation, share a common structure: expensive evaluations, high dimensional design spaces, and a need to combine data driven search with domain expertise. This project develops methodology for reliable AI assisted design along three complementary lines. The first is multi-fidelity Bayesian fusion, which combines cheap approximate simulations with expensive accurate ones to reduce the cost of exploring large design spaces. The second is Bayesian optimization with structured expert priors, which encodes domain knowledge as statistical priors rather than black box features so that optimization remains sample efficient and interpretable. The third is large language model based reasoning agents that propose candidate designs and verify them against domain constraints, aiming to make agentic search auditable rather than opaque. The methods will be validated primarily on analog and mixed signal circuit design, including transistor sizing, yield estimation, and simulation acceleration, an area with an established track record of peer reviewed results at DAC, ICCAD, and ASP-DAC. The underlying methodology is not specific to circuits and extends naturally to other engineering domains with similar cost and reliability constraints.