Reducing Uncertainties in Coupled Carbon‐Water Cycle Predictions—A Parameter Perturbation Ensemble Experiment at Three NEON Tower Sites in the Southeastern United States

Abstract A newly developed simulation tool, the National Center for Atmospheric Research–National Ecological Observatory Network (NCAR‐NEON) modeling framework, offers a promising opportunity to investigate parameter sensitivity in climate models. This is primarily due to its ultra‐efficient computational design and the availability of high‐quality NEON observational data. We conducted a Parameter Perturbation Ensemble (PPE) experiment using the Community Land Model version 5 (CLM5) at three NEON tower sites in the Southeastern U.S. We analyzed a subset of 30 parameters that influence carbon‐water cycle processes, running a total of 244,800 yrs of CLM5 simulations, which included 200 yrs of spin‐up and 4 yrs of production runs for each of the 400 parameter sets across three sites (3 sites × 400 parameter sets × 204 yrs). The best‐performing parameter set was identified by equally weighting normalized model performance for gross primary productivity (GPP) and evapotranspiration (ET). Joint calibration of both GPP and ET achieved the highest overall model performance (0.89), outperforming GPP‐only (0.75), ET‐only (0.78), and the default configuration (0.70) on a normalized scale from 0 to 1. Additionally, optimization produced a revised value for the hydraulic trait parameter psi50 that aligns more closely with trait‐based observations. When applied regionally at Ozarks Complex domain, the optimized point‐scale parameters reduced gross primary productivity overestimation by 39% and evapotranspiration underestimation by 10% compared to the default CLM5 configuration. Despite these improvements, challenges in upscaling remain due to uncertainties in input data and reference data sets, as well as the limited temporal and spatial coverage of NEON observations.