Abstract: Ecosystem temporal stability (TS) determines its ability to maintain structure, function and services under external disturbances, playing a critical role in the global carbon cycle and climate regulation. However, the capability of numerical models to simulate the TS of ecosystem carbon uptake remains insufficiently assessed. This study evaluated the performance of nine terrestrial ecosystem models in simulating gross primary productivity (GPP) and its TS and employed Random Forest (RF) models with Shapley Additive Explanations (SHAP) to identify key factors contributing to model biases. Site-scale analysis based on flux tower observations indicated that most models underestimated GPP while overestimating its TS, with the most pronounced biases occurring at the interannual scale. These discrepancies primarily stemmed from errors in simulating vegetation phenology, specifically the carbon uptake period and physiological traits, particularly peak GPP within a year. At the global scale, regions with higher carbon uptake tended to exhibit greater TS, yet significant discrepancies existed among models. Notably, RF and SHAP analyses indicated that leaf area index was more important than climate and geographical factors in explaining model divergence for simulating GPP and its TS. The study revealed systematic biases in the current models’ representation of TS, highlighting the potential vulnerability of ecosystems. These uncertainties among models may lead to an overestimation of ecosystem resilience, introducing uncertainties in global carbon budget estimates and potentially misguiding scientific assessments and policy decisions regarding future climate change responses. Therefore, improving carbon cycle simulation mechanisms is essential for enhancing model predictive capabilities.

Key words: terrestrial ecosystem models, gross primary productivity, climate change, carbon sinks, ecosystem stability, remote sensing