Benchmarking the seasonal cycle of CO₂ fluxes simulated by terrestrial ecosystem models

We evaluated the seasonality of CO₂ fluxes simulated by nine terrestrial ecosystem models of the TRENDY project against (1) the seasonal cycle of gross primary production (GPP) and net ecosystem exchange (NEE) measured at flux tower sites over different biomes, (2) gridded monthly Model Tree Ensembles-estimated GPP (MTE-GPP) and MTE-NEE obtained by interpolating many flux tower measurements with a machine-learning algorithm, (3) atmospheric CO₂ mole fraction measurements at surface sites, and (4) CO₂ total columns (X_CO2) measurements from the Total Carbon Column Observing Network (TCCON). For comparison with atmospheric CO₂ measurements, the LMDZ4 transport model was run with time-varying CO₂ fluxes of each model as surface boundary conditions. Seven out of the nine models overestimate the seasonal amplitude of GPP and produce a too early start in spring at most flux sites. Despite their positive bias for GPP, the nine models underestimate NEE at most flux sites and in the Northern Hemisphere compared with MTE-NEE. Comparison with surface atmospheric CO₂ measurements confirms that most models underestimate the seasonal amplitude of NEE in the Northern Hemisphere (except CLM4C and SDGVM). Comparison with TCCON data also shows that the seasonal amplitude of X_CO2 is underestimated by more than 10% for seven out of the nine models (except for CLM4C and SDGVM) and that the MTE-NEE product is closer to the TCCON data using LMDZ4. From CO₂ columns measured routinely at 10 TCCON sites, the constrained amplitude of NEE over the Northern Hemisphere is of 1.6 ± 0.4 gC m^-2 d^-1, which translates into a net CO₂ uptake during the carbon uptake period in the Northern Hemisphere of 7.9 ± 2.0 PgC yr^-1.