Land surface models, when connected to vegetation demographic models, are essential tools for predicting the fate of carbon and carbon-water-energy feedbacks in the Earth system, as well as for predicting current and future vegetation dynamics and their interactions with their environment. These models often contain many parameters, some of which represent measurable quantities (albeit with high uncertainty), and others which have no real-world analog. Quantifying and constraining parametric uncertainty is therefore a critical and yet difficult problem to overcome before such models can be relied upon to robustly simulate land surface dynamics and vegetation demography. Such exercises are difficult due to the nonlinear nature of parameter interactions as well as the high computational cost of running large ensembles.

The Functionally Assembled Terrestrial Ecosystem Simulator (FATES) is a vegetation demography model that can be run alongside a host land surface model to predict global vegetation dynamics. Importantly, FATES can be run in various complexity modes, where different processes within the model are driven by observations, allowing for better calibration of specific model parameters and structural components. Parameters tuned in a simpler mode can be then held fixed as new parameters are calibrated in more complex modes.

In this study, we train machine learning-based emulators to predict gross primary production, evapotranspiration, and other energy fluxes in FATES satellite phenology mode (FATES-SP, where leaf area index and plant height are driven by satellite observations) as a function of FATES input parameters. These emulators are then used to fully explore the FATES-SP parameter space and identify parameter sets that yield simulations within observational uncertainty. Parameter sets that minimize model bias are used to generate a model ensemble that explores uncertainty in photosynthesis and land-surface energy exchange of the calibrated model under future climate scenarios. Successful parameter calibration in FATES-SP mode will allow for increased confidence in current and future predictions of SP mode simulations, and will provide a starting point for calibration in increased complexity modes.