Overestimated temperature sensitivity in future net primary productivity projections

Terrestrial net primary productivity (NPP) is a fundamental indicator of ecosystem functioning and carbon sequestration capacity, playing a crucial role in regulating the strength of the land carbon sink. Earth system models (ESMs) are the primary tools for simulating historical NPP variations and projecting future changes. However, NPP projections by ESMs remain subject to substantial uncertainty and bias due to differences in model structure and parameterization. Here, we develop an observation-constrained transfer learning framework that integrates 21 CMIP6 ESMs with satellite-derived NPP to reduce structural biases and constrain future projections. The results indicate that the framework effectively reduces the mean absolute historical bias between ESM simulations and satellite-derived observations by 75.7%. Raw ESM projections show increases in global NPP of 85.88 ± 41.90 and 121.83 ± 67.09 gC m −2 yr −1 under SSP245 and SSP585, respectively. In contrast, observation-constrained projections reduce these increases to 55.23 ± 9.59 and 89.51 ± 18.46 gC m −2 yr −1 , while narrowing the 5th–95th percentile uncertainty range by 78%–82%. Factorial attribution shows that temperature dominates the projected NPP trend among the climatic drivers considered, while precipitation and interaction effects are comparatively small. Temperature-sensitivity analyses further indicate stronger NPP responses to warming in unconstrained ESMs, particularly in tropical regions, whereas observational constraints moderate these responses and lead to lower future NPP projections. Overall, our results suggest that the stronger temperature sensitivity identified in unconstrained ESMs is associated with larger projected increases in terrestrial NPP.