High-resolution GHG flux inversion by emulating atmospheric transport using machine learning.

Previous work has shown that point sources are responsible for a large percentage of the Greenhouse gases (GHGs) emission budget and studying them necessitates densely spaced measurements due to their localized nature. Constructing the source-receptor operator (i.e., the “footprint”) allows researchers to conduct emission flux inversion. The construction of footprints often becomes computationally intractable as the number of measurements increases, as is the case with the next-generation dense observing systems. Here we use FootNet which is a deep-learning-based model and serves as a 1000x faster surrogate for the full-physics-based atmospheric transport model. The FootNet contains U-Net architecture and is trained on footprints generated by the Weather Research & Forecasting – Stochastic Time-Inverted Lagrangian Transport (WRF-STILT) model. We used the footprints generated by the FootNet to compute hourly urban CO2 emission fluxes in the Bay area using the Bayesian Inversion method. The emission fluxes computed with FootNet agree very well with a case study from the published literature. We show the feasibility of using FootNet for near-real-time construction of footprints and computation of emission fluxes. Going forward, this will allow us to efficiently interpret dense GHG measurements in near-real-time at one-kilometer scale resolution.