Ying-Ping Wang: Applications of machine learning to improve global soil carbon predictions for Earth system models
Soils store as much carbon as vegetation and the atmosphere combined. Even small changes in soil carbon stocks can therefore have major implications for carbon–climate feedbacks and future climate projections. However, current Earth system models explain less than 20% of the observed global variation in soil carbon stocks, limiting confidence in their long-term projections.
In this talk, we show how machine learning can be used to improve the representation of poorly understood controls on soil carbon stabilisation, particularly the interacting effects of soil texture, pH and metal oxides on mineral-associated organic carbon. Using K-means cluster analysis, we classified global soils into 10 clusters based on harmonised datasets of clay and silt fractions, pH and metal oxide concentrations. For each cluster, we optimised key parameters in a newly developed next-generation soil carbon model that explicitly represents aggregation and organo-mineral interactions.
Our results show that combining machine learning with process-based modelling substantially improves global soil carbon prediction, increasing the explained variance in observed soil carbon stocks to approximately 50%. This approach provides a practical pathway for incorporating unresolved soil biogeochemical controls into land surface models. The modelling framework has been developed and tested for three widely used global land surface models, including CABLE, and is designed to support improved soil carbon representation in Earth system models.