Enhanced soil carbon dynamics in the Millennial model with data assimilation

Carbon (C) storage in soil, coupled with concerns regarding the impact of a warming climate on its stability, has elevated soil C into a global scientific and political discourse. The Millennial model is a next-generation soil C model that reflects our recent advancements in understanding soil C dynamics, including microbial decomposition, mineral association and aggregation. However, the model's simulations and predictions remain largely uncertain due to a lack of data, including the various soil C fractions, the complex model structure and many parameters. Despite recent progress, the Millennial model has not been well-tested against measurements related to the modelled C states. With the ever-increasing availability of high-quality spatially explicit data on climate, vegetation, and soil properties (e.g. from proximal and remote sensing), there is a need to integrate these to constrain the model and improve simulations and predictions. This research aims to reduce uncertainties using data assimilation. We first reduce the Millennial model uncertainty by updating the empirical equation of maximum sorption capacity with a more realistic estimation method. The measured soil C fractions and spatially explicit forcing inputs (e.g. NPP, soil moisture, soil temperature) across different ecosystems are used to reduce the observation and forcing uncertainty. We use multiple objective global sensitivity analyses to identify influential parameters and calibrate the parameters with an efficient parameter optimisation algorithm to reduce parameter uncertainty. The site-by-site optimisation method with measured C fractions was accurate, with an RMSE of 0.2 kg C/m² and a ρ_c of 0.97 for total organic carbon (TOC). Our future simulations by the Millennial model for the median changes of TOC in 2070--2100 across rangelands are 1.51 t/ha under SSP126 and 1.93 t/ha under SSP585, which are more conservative than the predicted changes by calibrated Roth-C model.