Global Methane Flux Inference using Emulated Atmospheric Transport

The Global Methane Pledge, signed at the United Nations Framework Convention on Climate Change 26th Conference of Parties, acknowledged the urgent need to rapidly cut methane emissions to meet climate change mitigation targets. However, atmospheric methane concentrations grew at the highest rate yet observed in 2021. Researchers have struggled to understand the relative contributions of natural and anthropogenic sources to the global methane budget, and there is intense debate about the causes of recent growth rate fluctuations. This uncertainty means that emission reduction efforts may not be optimally targeted, and the role of methane sources in climate change feedback is not well understood. Therefore, improved quantification of the global methane budget has been recognised as a grand challenge in environmental science for the coming decade (e.g., Ganesan et al., 2019). To support policymakers and researchers, millions of dollars are being spent annually on atmospheric methane observations, including on recently launched or planned satellites. Because of the enormous data volumes these systems are generating, they have the potential to revolutionise our ability to evaluate climate agreements and understand methane-related climate feedbacks. To meet this potential, a step-change is now needed in our methods for inferring fluxes from very large atmospheric datasets. Our proposal will develop a greenhouse gas flux inference system using new developments in machine learning and hierarchical Bayesian inference. Our approach builds on two recent proof-of-concept studies in these fields, led by our group. Firstly, we will develop a machine learning emulator of the transport of greenhouse gases through the atmosphere. This method allows the "footprint" of a greenhouse gas observation to be directly estimated from meteorological fields, in a fraction of a second. We aim to use new deep learning technologies such as graph neural networks to emulate greenhouse gas transport and dispersion at any location on the planet, opening the possibility of applying our algorithms to satellite observations. Secondly, we will develop a highly efficient method to solve for greenhouse gas emissions using data and models in a hierarchical Bayesian framework. This hierarchical approach, which can be shown to lead to more robust uncertainty estimates than traditional Bayesian methods, was demonstrated by our team for regional emissions evaluation using in situ measurement data. Here, we will further extend it for global flux inference. Based on these advances, we will constrain global methane emissions using satellite data between 2019 and 2027 and beyond, providing new information with which to inform policy and process understanding of methane fluxes.