Air pollution is one of the most pressing global challenges, with severe consequences for human health, ecosystems, and climate. Atmospheric pollutant dispersion is governed by turbulent
transport and diffusion, and mathematical models are used to simulate concentration �fields that represent these processes. However, the accuracy of such models typically depends on
input parameters whose values may not be known exactly, making the use of uncertainty quantification (UQ) and sensitivity analysis tools essential. Conventional global sensitivity analysis
(GSA) often collapses spatial information into scalar measures, limiting its ability to reveal how the influence of individual parameters varies across space. This thesis applies spatially resolved GSA to steady-state advection-diffusion systems for modelling pollutant dispersal. Specifi�cally, Sobol indices are used to characterise how parameter influence varies across space, with con�fidence intervals constructed at each stage of analysis to assess the reliability of the sensitivity estimates. The investigation progresses from one-dimensional test problems, which establish
a benchmark and con�rm the variance in wind speed as the dominant source of output variability, to two-dimensional models that capture more realistic dispersion. The models use the
Streamline Upwind Petrov-Galerkin (SUPG) �finite element method to ensure robust and reliable discretisation of the problem. In the 2D case, computational demands are mitigated by the use of arti�ficial neural network (ANN) surrogate models trained on numerical simulations, enabling efficient estimation of Sobol indices, which serve as the GSA measure in this work.
This approach is then applied to an urban air quality model formulated as a full PDE system.
The subsequent study demonstrates both the dominant role of wind speed and the regional signi�ficance of the stability exponent and emission rates, with ANNs again providing scalability while preserving key spatial sensitivity patterns. The fi�ndings show that spatially resolved GSA reveals parameter influence in ways inaccessible to scalar approaches, highlighting how sensitivities
shift across space and how targeted reduction of uncertainty in key parameters can lower output variance. By combining numerical modelling, surrogate methods, and UQ, the thesis demonstrates that high-resolution spatial GSA is both feasible and informative for pollutant dispersal problems, and proposes a transferable methodology for other PDE-based models in which spatial variability is key.
| Date of Award | 13 Feb 2026 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | EPSRC (Engineering and Physical Sciences Research Council) |
|---|
| Supervisor | John MacKenzie (Supervisor) & Alison Ramage (Supervisor) |
|---|