Greenhouse gas (GHG) emissions are a primary driver of contemporary climate change, contributing to rising global temperatures, increasing frequency of extreme weather events, and widespread ecological disruption. Effective mitigation depends not only on reducing emissions but also on accurately detecting, locating and quantifying them. Reliable source characterisation underpins national climate strategies, industrial compliance, and international agreements aimed at stabilising the global climate system. This thesis develops probabilistic inversion frameworks that integrate atmospheric trans port models, Bayesian inference, and machine learning to improve the estimation of gas emission source characteristics from ground-based measurements. First, we address limitations of the widely used Gaussian plume model, where dispersion parameters are often fixed via atmospheric stability classes, introducing bias when meteorological classifications are inaccurate. We propose a gradient-based Markov chain Monte Carlo inversion scheme that jointly infers dispersion parameters alongside source location, emission rate, background concentration, and sensor error. Application to both controlled-release data and simulations demonstrates improved accuracy and uncertainty quantification compared to traditional methods. Second, we tackle the challenge of real-time inversion in obstructed, unsteady-state flow fields, where computational fluid dynamics (CFD) solvers are too expensive for sequential inference. We design deep-learning surrogate models trained on high-fidelity CFD outputs and embed them within particle filters, enabling near-instantaneous Bayesian estimation of time-varying source parameters. Validation on the Chilbolton methane release dataset and complex synthetic environments shows comparable accuracy to full CFD inversion at orders-of-magnitude lower computational cost. Finally, we synthesise these contributions, explore integration with emerging satellite based inversion systems, and outline pathways for scaling these methods to regional and global monitoring networks. Together, these contributions provide physically grounded, computationally efficient tools for GHG monitoring. By advancing both parameter estimation accuracy and operational feasibility, this work supports scalable, uncertainty-aware frameworks for emissions quantification, informing policy, compliance, and mitigation strategies.
Thomas B. Newman (Thu,) studied this question.