Atmospheric inversions have been used to assess biosphere–atmosphere CO2 surface exchanges at various scales, but variability among inverse flux estimates remains significant, especially at continental scales. Atmospheric transport errors are one of the main contributors to this variability. To characterize transport errors and their spatiotemporal structures, we present an objective method to generate a calibrated ensemble adjusted with meteorological measurements collected across a region, here the upper US Midwest in midsummer. Using multiple model configurations of the Weather Research and Forecasting (WRF) model, we show that a reduced number of simulations (less than 10 members) reproduces the transport error characteristics of a 45-member ensemble while minimizing the size of the ensemble. The large ensemble of 45 members was constructed using different physics parameterization (i.e., land surface models (LSMs), planetary boundary layer (PBL) schemes, cumulus parameterizations and microphysics parameterizations) and meteorological initial/boundary conditions. All the different models were coupled to CO2 fluxes and lateral boundary conditions from CarbonTracker to simulate CO2 mole fractions. Observed meteorological variables critical to inverse flux estimates, PBL wind speed, PBL wind direction and PBL height are used to calibrate our ensemble over the region. Two optimization techniques (i.e., simulated annealing and a genetic algorithm) are used for the selection of the optimal ensemble using the flatness of the rank histograms as the main criterion. We also choose model configurations that minimize the systematic errors (i.e., monthly biases) in the ensemble. We evaluate the impact of transport errors on atmospheric CO2 mole fraction to represent up to 40 % of the model–data mismatch (fraction of the total variance). We conclude that a carefully chosen subset of the physics ensemble can represent the uncertainties in the full ensemble, and that transport ensembles calibrated with relevant meteorological variables provide a promising path forward for improving the treatment of transport uncertainties in atmospheric inverse flux estimates.