PhD: Diffusion-Based Generative AI and Bayesian Optimisation for Water Pollution in Complex Domains and Networks

University of Glasgow

Glasgow, UK 🇬🇧

Project institution: University of Glasgow

Project supervisor(s): Dr Mu Niu (University of Glasgow), Dr Craig Wilkie (University of Glasgow), Prof Cathy Yi-Hsuan Chen (University of Glasgow), Dr Michael Tso (UK Centre for Ecology & Hydrology)

Research area: Atmosphere, Hydrosphere, Cryosphere, and Ecosystem Processes and Evolution

​​Technological advances in smart sensors are changing how the environment can be monitored. Small, energy-efficient and mobile sensors provide new opportunities for water-quality monitoring, but their effectiveness depends critically on where sensors are deployed and how sparse observations are used to infer pollution across the wider water system. 

​This PhD project aims to develop a mathematical and computational framework combining diffusion-based generative AI, Gaussian processes (GPs), and Bayesian optimisation (BO) for modelling and monitoring water pollution over complex constrained domains and networks.​ 

Methods Used:

​This project will develop a mathematical and computational framework combining diffusion-based generative AI, and intrinsic Bayesian optimisation (BO) for modelling and monitoring water pollution over complex constrained domains and networks. 

​Water-quality measurements are often available at only a small number of sensor locations. The project will investigate conditional diffusion-based generative models to reconstruct complete pollution fields from these sparse observations. Observed sensor measurements provide partial information about the pollution field, while the generative model produces plausible values over unobserved regions. Multiple generated pollution fields can represent different scenarios consistent with the available observations, providing a flexible way to quantify uncertainty. 

​A major methodological challenge is that water systems often have complex non-Euclidean structures. Lakes can have irregular boundaries and complex internal geometry, while drainage and sewage systems naturally form networks. Standard generative models designed for regular Euclidean grids may fail to represent these structures appropriately. The project will therefore investigate geometry-aware diffusion models that incorporate the intrinsic geometry and connectivity of the water system, using tools such as heat kernels, graph Laplacians and diffusion processes on non-Euclidean spaces. 

​The project will also develop intrinsic Bayesian optimisation for adaptive sensor placement. Intrinsic Gaussian Processes on constrained domains and graphs will provide predictions and uncertainty estimates that respect the geometry of the water system. BO will use this information to identify informative locations for additional measurements. The generative and BO components will subsequently be integrated, allowing uncertainty obtained from generated pollution fields to inform where new sensors should be deployed. 

​The methodology will be extended to spatio-temporal settings to account for changing pollution patterns over time. Scalable computation and GPU parallelisation will also be explored for large environmental datasets. 

​Teaser Project 1 Objectives 

​The first teaser project will focus on intrinsic Gaussian processes and Bayesian optimisation for water-quality monitoring. The student will apply existing GP methods on constrained domains and networks to small-scale water-quality datasets and investigate how domain geometry and network connectivity influence prediction and uncertainty quantification. 

​The student will then explore Bayesian optimisation and adaptive sampling, using GP predictions and uncertainty to determine informative locations for additional water-quality measurements. This project will provide foundations in Gaussian processes, non-Euclidean statistical modelling and Bayesian optimisation, and can subsequently develop into new intrinsic BO methods for complex domains and networks. 

​The student will develop GPU-accelerated implementations of intrinsic GP inference and Bayesian optimisation, using GPU computing frameworks such as PyTorch. This will include parallelising computationally intensive matrix operations, graph-based covariance calculations and the evaluation of candidate sensor locations. The student will benchmark GPU implementations against CPU-based methods and investigate their scalability to larger spatial domain. 

​Teaser Project 2 Objectives 

​The second teaser project will focus on diffusion-based generative AI for water-quality modelling. The student will first investigate existing diffusion and score-based generative models and their use for reconstruction from sparse observations. 

​The student will apply these ideas to small-scale water-quality datasets, treating sensor measurements as observed parts of an otherwise unknown pollution field. Diffusion models will be used to generate plausible complete pollution fields conditional on the available measurements and to investigate uncertainty in unobserved regions. 

​The student will develop GPU-accelerated implementations of diffusion model training and sampling using frameworks such as PyTorch. Particular attention will be given to parallelising stochastic diffusion simulations, neural-network computations. The student will evaluate computational performance and investigate scalable implementations for larger spatial domains.  

​This work can subsequently develop into geometry-aware diffusion models for constrained domains and networks, with extensions to spatio-temporal pollution modelling. Ultimately, these generative models can be integrated with intrinsic Bayesian optimisation so that uncertainty about the pollution field guides the adaptive placement of new sensors.

69 days remaining

Apply by 31 January, 2027

POSITION TYPE

ORGANIZATION TYPE

EXPERIENCE-LEVEL

DEGREE REQUIRED

Register for World Toilet Summit 2026 - Thailand - November 17-19, 2026