GATES-Background 0.2.0: Emulating background greenhouse gas mole fractions for regional atmospheric inverse modelling with graph neural networks

Abstract. Regional atmospheric trace gas inverse modelling frameworks depend on accurately simulating mole fractions, which result from transporting surface fluxes within a domain and carrying mole fractions into the region from its boundaries. With the aim of improving the computational efficiency of inverse modelling pipelines, recent machine learning approaches have successfully emulated Lagrangian Particle Dispersion Model (LPDM) footprints (source-receptor relationships) at substantially reduced computational cost, compared to the numerical model simulations. However, the estimation of the background component of the simulated mole fractions, an important contributor to overall mole fraction variability, has received comparatively little attention. In this study, we build on the GATES model (Graph-Neural-Network Atmospheric Transport Emulation System, Fillola et al. (2026b)), which emulates LPDM outputs, and present GATES-Background 0.2.0, a complementary emulator for background mole fraction contributions to observations across a regional domain driven by meteorological data and large-scale background estimates from a global methane reanalysis.

GATES-Background is trained and evaluated to emulate background concentrations for GOSAT observations over South America and demonstrates strong agreement with reference calculations generated from the propagation of boundary conditions to the measurement location. The emulator generally captures daily variability, seasonal structure, and regional gradients in background mole fraction contributions, with errors that are relatively small and spatially localized, compared to the variability in the simulated background values (root-mean-square errors of 5.3–7.4 ppb, depending on season, compared to GOSAT XCH4 retrieval error on the order of 13 ppb and estimated variability due to boundary conditions within this region of approximately 15–20 ppb). The proposed approach reduces computational cost by several orders of magnitude relative to physics-based background calculations, enabling scalable and near–real-time application. When combined with GATES or other LPDM footprint emulators, this work will enable efficient machine learning–based approximation of both components of the forward model used in regional LPDM-based greenhouse gas inversion systems.