Why mechanistic models perform poorly in global soil organic carbon prediction

State-of-the-art mechanistic models perform poorly to accurately capture the amount and spatial variability of global soil organic carbon (SOC) stocks compared to machine learning models. Identifying the reasons for these shortcomings using interpretable machine learning techniques is essential to advancing our understanding of SOC turnover processes and guiding future model development. In this study, we trained both mechanistic and machine learning models using approximately 37,000 global SOC observations. The machine learning models consistently outperformed the mechanistic models, achieving higher R² values and lower RMSE. To diagnose the limitations of mechanistic models, we trained random forest models with the mechanistic model inputs as predictors and either observed or modelled SOC as the target variable. Applying multiple explainable artificial intelligence (XAI) techniques including feature importance, partial dependence plots, and SHapley Additive exPlanations (SHAP), we found that while the trends in SOC responses to environmental variables were comparable between observed and modelled SOC, the magnitude of SOC sensitivity to different variables in the mechanistic models was weaker. Furthermore, the distribution of partial dependence values for observed SOC across specific variables was poorly represented by mechanistic models, even though the mean partial dependence values were similar. Notably, soil moisture and pH were significantly under-represented in the mechanistic models, highlighting the need for further research on the dependence of SOC turnover on these variables. Our study showed that XAI techniques can effectively reveal how well individual variables and their combined effects are represented in the mechanistic models, providing clear and specific direction for future model development.