As atmospheric levels of carbon dioxide levels continue to increase, it is critical that terrestrial ecosystem models can accurately predict ecological responses to the changing environment. Current predictions of net primary productivity (NPP) in response to elevated atmospheric CO2 concentration and/or elevated temperatures are highly variable and contain a considerable amount of uncertainty.
Benchmarking these model predictions against data are necessary, not only to assess individual models on their ability to replicate observed patterns, but also to identify and evaluate the assumptions causing inter-model differences. However, to keep up with the rate at which models are being developed, the ecosystem modeling community needs to be able to perform large scale model intercomparisons at a higher frequency than previously possible.
The Predictive Ecosystem Analyzer (PEcAn) is an informatics toolbox that wraps around an ecosystem model and can be used to help identify which factors drive uncertainty. We used PEcAn to perform a model intercomparison using models that represent a range from low to high structural complexity, across a range of sites from the Free-Air CO2 Enrichment (FACE) experiments. These studies provide us with a wide range of ecosystem properties at scales that are directly comparable to models.
Model assessment tests were implemented in a novel benchmarking workflow that is automated, repeatable, and generalized to incorporate different sites and ecological models. Observational data from the climate manipulation experiments represent a first test of this flexible, extensible approach aimed at providing repeatable tests of model process representation that can be performed quickly and frequently.
Combining the observed patterns of uncertainty between multiple models with results of the recent FACE-model data synthesis project (FACE-MDS) helped identify which processes need further study and additional data constraints. While many models produced similar estimates of NPP, they differed greatly in their predictions of the components of NPP such as leaf area index and biomass, exposing processes in which models may have been getting the right answers for the wrong reasons. These findings can be used to inform future experimental design and in turn can provide informative starting point for data assimilation.