Abstract: Methane (CH₄) is the second most important greenhouse gas after carbon dioxide, accounting for 16-25% of atmospheric warming to date. However, large uncertainty exists in methane emissions estimates using biogeochemistry models. This uncertainty arises largely because CH₄ dynamics depend on multiple physical, biological, and chemical processes and a large number of uncertain model parameters. Sensitivity analysis (SA) can help not only identify important parameters for methane emission, but also achieve reduced biases and uncertainties in future projections.

In this study, SA is performed for the pre-selected critical parameters of methane biogeochemistry module within the Energy Exascale Earth System Model (E3SM) land model (ELM). Considering the large number of model simulations typically required for the variance-based SA, we employ a machine learning algorithm, namely, Gaussian process regression, to construct a surrogate model that enables emulating the behaviour of ELM methane biogeochemistry and conducting a full variance-based SA with much reduced computational costs but barely any loss in accuracy.

We examine the sensitivity of CH₄ emission to a large set of parameters at multiple FLUXNET-CH4 sites of different vegetation types. Our results will provide useful insights into the key parameters and processes that drive the uncertainty in methane emissions at different sites and the consistency of parametric uncertainties across vegetation types, soil types, climatic zones, and seasons.