Accelerating Greenhouse Gas Emissions Inference using Machine Learning

: Developing Emulators for Lagrangian Atmospheric Transport

Abstract

There is an urgent need for greenhouse gas (GHG) monitoring frameworks that are robust, near-real-time and high-resolution. Inverse methods use atmospheric measurements from sites or satellites to estimate surface emissions (“fluxes”), providing independent verification to complement traditional national reporting. However, many well-established inverse modelling techniques face computational challenges scaling to the growing volume of observational datasets, particularly those relying on Lagrangian Particle Dispersion Models (LPDM) to simulate atmospheric transport. These models trace the movement of ‘air parcels’ backwards from the measurement location to identify the surface regions influencing that observation, producing a two-dimensional sensitivity map known as ‘footprint’.
This work explores Machine Learning (ML) approaches to emulate LPDMs, with the goal of outputting footprints from meteorological inputs orders of magnitude faster than traditional methods, while maintaining comparable accuracy. A proof-of-concept emulator is first developed for seven measurement sites worldwide. The emulated footprints are evaluated and integrated into a Hierarchical Bayesian inversion framework to estimate methane emissions over the UK in 2016, producing fluxes consistent with those derived from full LPDM simulations.
The focus then turns to LPDM-driven satellite inversions. The Graph-Neural Network Atmospheric Transport Emulation System (GATES) model is introduced, adapting novel Graph-Neural-Network (GNN) architectures for data-driven weather forecasting to atmospheric transport. The model is applied to two case studies using GOSAT satellite observations: one over Brazil and another over North Africa. Across both regions, GATES accurately reproduces LPDM footprints while achieving a ×1000 speed-up (1 second per footprint versus 20 minutes with the LPDM). The emulated satellite footprints are also integrated within inversions to estimate methane emissions across multiple countries, producing results that are spatially and temporally consistent with LPDM-driven inversions and recent literature estimates.
Building on these results, this thesis investigates the potential of transfer learning to extend GATES beyond regional applications. Joint training and fine-tuning experiments evaluate how knowledge learned in one region can be adapted to the other. The models are shown to learn region-specific transport, but the fine-tuning experiments indicate the possibility of transferring knowledge across regions. Finally, the limitations of the model are discussed, including the influence of data sparsity and regional variability on model errors, along with lessons learned and proposed directions for future work. Overall, these findings represent a first step towards scalable, data-driven atmospheric transport models capable of supporting near-real-time global inversions.

Date of Award	21 May 2026
Original language	English
Awarding Institution	University of Bristol
Supervisor	Matthew L Rigby (Supervisor) & Raul Santos-Rodriguez (Supervisor)