Disentangling Sources of Uncertainty in CLM5 Model Predictions: Water, Energy, and Carbon Fluxes at European Observation Sites

Abstract

Land surface models (LSMs) such as the Community Land Model version 5 (CLM5) are central to climate–carbon assessments but exhibit persistent biases relative to site‐level observations. It remains unclear to what extent model‐data mismatches arise from uncertainty in meteorological forcing, soil properties, or vegetation traits, and to what extent they instead reflect structural deficiencies. We evaluated surface and root‐zone soil moisture (SM, SMr), evapotranspiration (ET), sensible heat flux (H), net ecosystem exchange (NEE), and gross primary production (GPP) at 14 European eddy‐covariance sites using 128‐member CLM5 ensembles in which (a) atmospheric inputs, (b) soil parameters, and (c) vegetation parameters were perturbed. Soil parameters dominate ensemble spread in SM/SMr, whereas vegetation parameters dominate spread in ET, H, NEE, and GPP. Despite this, substantial mean biases remained (0.18 cm 3  cm −3 for SM, −0.22 mm d −1 for ET, −19.7 W m −2 for H, −1.83 gC m −2  d −1 for GPP, and 1.16 gC m −2  d −1 for NEE), and the 99% ensemble envelopes failed to cover many observations. Coverage percentage (CP; fraction of observations within the 99% ensemble envelope) was 74.3% (SM), 74.6% (SMr), 68.8% (ET), 47.7% (H), 65.1% (NEE), and 70.1% (GPP). A companion CLM5‐SP (Satellite Phenology) experiment with prescribed ICOS leaf area index (LAI) confirmed that prognostic LAI errors in CLM5‐BGC (Biogeochemistry) are the primary structural driver of SM/SMr and GPP coverage deficits, while BGC's carbon–nitrogen coupling better captures ET and H variability. CP diagnostics delineate parametric from structural error regimes across variables and plant functional types.