Emulating Atmospheric Transport using Machine Learning for Greenhouse Gas Emission Flux Estimation

Abstract

Carbon dioxide and methane are the two strongest anthropogenic greenhouse gases (GHGs) and together they account for more than 85% of the GHG radiative forcing since pre-industrial times. The future states of our climate have a profound impact due to their past, current and future emissions. Quantifying their emissions is important to understand why the global concentrations of GHGs are rising. Densely spaced measurements are required to study the emissions from the point sources which are responsible for a large percentage of the total emission budget. Estimating GHG emissions using atmospheric measurements is typically done by constructing source-receptor relationships (also known as “footprints”). Constructing these footprints using full-physics atmospheric transport models (ATMs) can be computationally expensive while working with densely spaced measurements at high spatio-temporal resolution. The outputs of these ATMs are storage expensive as well. Here we developed FootNet, a deep-learning emulator for atmospheric transport at a kilometer scale. The emulator is trained and evaluated using footprints simulated using a Lagrangian Particle Dispersion Model (LPDM). This emulator is completely independent of any full-physics atmospheric transport model and only uses meteorological parameters as inputs. The emulator predicts the magnitude and spatial pattern of the footprints in near-real-time and hence addresses the computational bottlenecks of the GHG flux inversion frameworks.