Why do inverse models disagree? A case study with two European CO(2 )inversions

We present an analysis of atmospheric transport impact on estimating CO2 fluxes using two atmospheric inversion systems (CarboScope-Regional (CSR) and Lund University Modular Inversion Algorithm (LUMIA)) over Europe in 2018. The main focus of this study is to quantify the dominant drivers of spread amid CO2 estimates derived from atmospheric tracer inversions. The Lagrangian transport models STILT (Stochastic Time-Inverted Lagrangian Transport) and FLEXPART (FLEXible PARTicle) were used to assess the impact of mesoscale transport. The impact of lateral boundary conditions for CO2 was assessed by using two different estimates from the global inversion systems CarboScope (TM3) and TM5-4DVAR. CO2 estimates calculated with an ensemble of eight inversions differing in the regional and global transport models, as well as the inversion systems, show a relatively large spread for the annual fluxes, ranging between -0.72 and 0.20 PgC yr(-1), which is larger than the a priori uncertainty of 0.47 PgC yr(-1). The discrepancies in annual budget are primarily caused by differences in the mesoscale transport model (0.51 PgC yr(-1)), in comparison with 0.23 and 0.10 PgC yr(-1) that resulted from the far-field contributions and the inversion systems, respectively. Additionally, varying the mesoscale transport caused large discrepancies in spatial and temporal patterns, while changing the lateral boundary conditions led to more homogeneous spatial and temporal impact. We further investigated the origin of the discrepancies between transport models. The meteorological forcing parameters (forecasts versus reanalysis obtained from ECMWF data products) used to drive the transport models are responsible for a small part of the differences in CO2 estimates, but the largest impact seems to come from the transport model schemes. Although a good convergence in the differences between the inversion systems was achieved by applying a strict protocol of using identical prior fluxes and atmospheric datasets, there was a non-negligible impact arising from applying a different inversion system. Specifically, the choice of prior error structure accounted for a large part of system-to-system differences.