Using Gaussian process emulation to quantify the global methane budget

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)


The atmospheric concentration of methane (the second most important greenhouse gas) is increasing, despite plateauing between 2000 and 2007. These changes remain unexplained despite a multitude of studies, each suggesting different primary drivers.

This thesis works towards constraining the methane budget and thoroughly quantifying its uncertainty. Gaussian process emulators are developed, which estimate the relationship between uncertain inputs (methane sources and sinks) and observable outputs (monthly average hemispheric mole fraction and δ13C-CH4 time series) for a three-dimensional chemical transport model (MOZART), without assuming
linearity. The emulators run as fast as two-dimensional box models, but have interannually varying transport and MOZART’s spatial resolution. Some minor source and sink parameters were held constant, and their uncertainty was, for the first time, included by considering the range of observable outputs when these parameters are varied within their uncertainties. This invariant parameter uncertainty was found to be comparable to previously estimated model uncertainties and thus should be considered in future studies.

Using the emulators, two analyses that would have been unfeasible for the computational expense of MOZART were carried out. The first is a sensitivity analysis to find the methane sources and sinks whose uncertainties cause the largest amount of variance in the modelled observable outputs. This analysis showed several parameters (such as the freshwater source, the Cl loss, and the initial conditions) that have often had their uncertainty ignored previously, have large effects on the modelled observable outputs. In the second analysis, the emulator output was compared to atmospheric observations to reduce the uncertainty ranges of the sources and sinks. However, only 10 out of 28 parameters were constrained, and the uncertainty ranges remained large. These large parameter ranges could explain the disagreement between many previous studies, which came to different conclusions about recent methane changes, despite using the same atmospheric datasets.
Date of Award29 Sept 2020
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorSimon O'Doherty (Supervisor) & Matthew L Rigby (Supervisor)

Cite this