Tracing uncertainty in predicting peatland carbon responses to multiple warming and CO2 treatments in northern Minnesota

Yiqi Luo, Center for Ecosystem Science and Society (ECOSS), Northern Arizona University, AZ, Enqing Hou, Department of Biological Sciences, Northern Arizona University, Flagstaff, AZ, Daniel Ricciuto, Environmental Sciences Division and Climate Change Science Institute, Oak Ridge National Laboratory, Oak Ridge, TN and Paul J. Hanson, Climate Change Science Institute and Environmental Sciences Division, Oak Ridge National Laboratory, Oak Ridge, TN

Background/Question/Methods
Model intercomparison projects (MIPs) usually show a large uncertainty in predicting land carbon (C) dynamics. However, the underlying causes remain not well understood. To trace sources of model uncertainty, we developed a matrix-based MIP (Matrix MIP) by expressing eight land C cycle models (i.e. TEM, CENTURY, DALEC2, TECO, FDBC, CASA, CLM4.5, and ORCHIDEE) in a unified matrix form. The eight models differ greatly in complexity, with the number of C pools ranging from 2 in TEM to 100 in ORCHIDEE. Two models, CLM4.5 and ORCHIDEE have multiple soil layers while the other six models are not vertically resolved. All eight models were driven by the same gross primary production (GPP) and environmental variables (e.g., temperature and soil moisture) under ambient condition and in an elevated CO2 (900 ppm) treatment at the SPRUCE experimental site.
Results/Conclusions
Despite all models using the same GPP, the variability found among the models became increasingly larger from simulated net primary production (NPP) to net ecosystem production (NEP) to soil organic carbon (SOC) dynamics. Our transient traceability analysis revealed that NPP diverges among the models due to differences in plant C use efficiency; NEP due to combined differences in C residence times and NPP; SOC dynamics predominantly due to differences in baseline soil C residence times. Conversely, model-to-model uncertainty effectively disappears when we standardized matrices of environmental scalars, plant carbon partitioning coefficients, decomposition coefficients, and transfer coefficients one-by-one. Our study shows that model uncertainty in predicting land C dynamics can be analytically understood when all the land C models are expressed in a unified matrix form.