Terrestrial biosphere models (TBMs) are integral tools to study ecosystem–atmosphere carbon exchange. However, TBMs diverge markedly in their carbon flux estimates, limiting our ability to forecast climate change impacts on the terrestrial carbon cycle. Model evaluation has routinely focused on locally optimized processes and functional relationships, yet the space–time variability in carbon flux estimates at regional to continental scales has remained divergent as ever. Here, we leverage atmospheric CO₂ observations to explore emergent patterns in the divergence among TBM estimates of gross primary productivity (GPP) and net ecosystem exchange (NEE) over North America. To do so, we evaluate a suite of diagnostic, prognostic, and machine-learning TBMs and solar-induced fluorescence (SIF) data products based on how well their regional patterns explain the variability in biospheric CO₂ drawdown.

Models with GPP and NEE estimates that effectively reproduce atmospheric CO₂ variability (as is indicated by R² values) share a strong growing-season sink in the Midwest US croplands, whereas the remaining models tend to place most growing-season uptake in forests. The difference in model explanatory power depends mainly on how well models represent the seasonal cycle of the growing-season cropland sink, rather than the partition of fluxes across biomes. Our results suggest that improving model representation of cropland processes that govern the seasonality of fluxes, such as phenology and carbon allocation, is a priority for robust quantification of North American carbon exchange.