**General description**

FLEXINVERT-CO2 (FI-CO2) is a Bayesian inversion framework developed at NILU. Atmospheric transport is modelled using the Lagrangian model FLEXPART (version 10beta), which was run with ECMWF operational wind fields. The inversion is solved using the Lanczos conjugate gradient algorithm.

**Atmospheric transport set-up**

FLEXPART retro-plumes were calculated for each observation (using 10,000 virtual particles) and traced backwards in time for 5 days. Footprints were computed at hourly temporal and 0.5° x 0.5° spatial resolution. After 5 days, the particle positions on the global domain were used to calculate a weighting matrix, which was integrated with 3D initial concentration fields, to calculate an initial concentration for each observation. The background concentration for each observation was defined as the sum of the initial concentration (as defined above) plus the contribution of fluxes outside the inversion domain to the change in concentration. The initial concentration fields were calculated monthly using a bivariate interpolation of NOAA discrete sampling data and were interpolated in time to the termination time of each retro-plume for the calculation of each observation's initial concentration. (In future the initial concentration fields will be based on an assimilation product).

**Inversion set-up**

FI-CO2 computes increments of NEE for ecosystem regions. One NEE increment is computed for nighttime (local time 0-3 h) and one for daytime (local time 12-15h) for every 5-day period (average over 5 days). The night and day increments are then interpolated to the prior temporal resolution and added to the prior NEE estimate. Currently, ocean fluxes are not optimized.