An Adaptive Dynamic AI-driven ParameTerization (ADAPT) framework for global hourly gross primary productivity estimation

High-frequency estimation of gross primary productivity (GPP) is crucial for understanding terrestrial carbon dynamics and ecosystem responses to rapidly changing environmental conditions. However, accurate GPP estimation at short-term timescales remains challenging, as it requires accounting for both instantaneous physiological reactions and legacy effects from antecedent environments. While light response curve (LRC) models effectively represent diurnal photosynthetic variations by capturing nonlinear light saturation, their global application has been constrained by static parameterizations. To address this, we propose the Adaptive Dynamic AI-driven ParameTerization (ADAPT) framework, which couples an LRC model with a deep learning approach. By integrating satellite-derived vegetation indices and 15-day antecedent meteorological statistics, ADAPT dynamically retrieves daily physiological parameters, enabling spatially explicit and temporally dynamic LRC parameterization. Using 328 eddy covariance flux sites globally (2001–2024), the dynamically parameterized model (LRC-DP) substantially mitigated underestimation biases associated with static models, achieving a site-level root mean squared error (RMSE) of 343.7 g C m −2 yr −1 for annual GPP. Against existing hourly products, ADAPT reduced RMSE by 25.8–32.4%, successfully reproducing midday peak productivity by overcoming the limitations of linear light-use efficiency assumptions and purely data-driven approaches. Global application at 0.05° resolution demonstrated that ADAPT captures sub-daily photosynthetic sensitivities and interannual variability consistent with global products. Highlighting its robustness under extreme climate stress, ADAPT demonstrated enhanced capability over existing products by accurately capturing daily and sub-daily GPP declines during heatwaves in the United States and Australia. This study highlights that integrating physiological mechanisms with dynamic parameterizations effectively resolves traditional scaling bottlenecks, advancing our capability to monitor high-resolution land–atmosphere carbon exchanges under climate extremes.