Extended Data Fig. 6: For some models, NEE estimates do not outperform the same models’ GPP estimates in explaining observed atmospheric CO₂ variability.

Fractions of observed CO₂ variability explained by NEE estimates (\({R}^{2}_{{{{\rm{NEE}}}}}\)) vs. those explained by GPP estimates (\({R}^{2}_{{{{\rm{GPP}}}}}\)) from simulations from the MsTMIP v2 (pink squares), TRENDY v6 (brown circles), and FLUXCOM (orange diamonds) model ensembles for 2007–2010. A model that falls below the one-to-one dashed line has GPP estimates that explain observed atmospheric CO₂ variability better than that same model’s NEE estimates. Similar to Fig. 3, symbols that are half filled to the left (right) indicate models whose GPP (NEE) estimates explain a higher fraction of CO₂ variability than does incoming shortwave radiation (\({R}^{2}_{{{{\rm{SW}}}}}=0.23\); green dotted line). Data points that represent the same model participating in different ensembles (for example, LPJ-wsl in MsTMIP and TRENDY) are linked with thin lines.