Contrasting parametric sensitivities in two global vegetation models using parameter perturbation ensembles

Uncertainty in land model projections remains high, yet the distinct roles of parametric errors versus structural model choices are difficult to disentangle. We compared parametric sensitivities in the Community Land Model (CLM) version 6.0 and the Functionally Assembled Terrestrial Ecosystem Simulator (FATES) operating in satellite phenology mode to isolate canopy biophysics. Using a one-at-a-time (OAAT) perturbation strategy, we modified over 300 parameters to their physical limits to quantify their effects on biophysical fluxes globally and across biomes. The resulting dataset contains global, zonal, and biome-level aggregations of these ensembles, alongside annual spatial maps of model output for each parameter perturbation. We found that while most parameters had minimal impact, the models exhibited contrasting sensitivities; CLM-FATES showed a larger spread in gross primary productivity (GPP) driven strongly by carboxylation rate , whereas CLM sensitivities were more distributed among vegetation and hydrology parameters. Additionally, CLM-FATES displayed higher water use efficiency and a dampened response to soil hydrology parameters compared to CLM. These divergence points underscore how model structure fundamentally alters parametric sensitivity. This dataset provides a comprehensive resource for identifying influential parameters to guide future calibration efforts and for investigating the mechanistic drivers of uncertainty in global land surface models.

This dataset contains post-processed model output from two global parameter perturbation ensembles (PPEs) generated using the Community Land Model (CLM) version 6.0. The ensembles are designed to compare parametric sensitivity and structural uncertainty between two vegetation model configurations: 1) the default CLM vegetation module, and 2) the Functionally Assembled Terrestrial Ecosystem Simulator (FATES). Both models were run in "satellite phenology" (SP) mode, where vegetation structure (LAI, PFT distribution) is prescribed via remote sensing data to isolate canopy biophysics and hydrological processes.

We employed a one-at-a-time (OAAT) perturbation strategy, modifying parameters independently to their estimated minimum and maximum physical bounds. Simulations were conducted on a 400-member "sparse grid" representative of global climatological and ecological heterogeneity, driven by GSWP3v2 reanalysis climate forcing cycled from 2000 to 2014.


File formats include several netcdf files that include global annual averages, by-biome annual averages, zonal (by-latitude) annual averages, and global monthly averages for several different experiments.

These data were generated and processed using Python (3.11) and the xarray (2024.1.1), pandas (2.2.0), numpy (1.24.3), and dask (2024.1.0) libraries. The workflow used to transform raw model history files into this post-processed dataset is documented in the associated GitHub repository: https://github.com/adrifoster/fates_calibration_library/. The Jupyter Notebooks provided in the GitHub repository (specifically under notebooks/JAMES_2025_OAAT_Manuscript) contain the exact logic used to generate the figures and tables in the accompanying JAMES manuscript from the data files provided in this repository.