Interpretable Machine Learning Reveals Global Nonlinear Responses of Soil Respiration to Climate

Abstract Soil respiration ( R s ) is a major component of the global carbon cycle, yet its responses to environmental drivers remain poorly constrained. Using 8,111 observations and a SHAP‐interpreted Random Forest model, we identified global response patterns for key environmental drivers of R s . Temperature contributions showed a peaked association with a descriptive maximum of 27.7°C, precipitation contributions plateaued at high values, and leaf area index contributions increased approximately linearly. These results provide a global, data‐driven characterization of peaked temperature and saturating precipitation associations, including how their shapes and supported ranges vary among climate zones. Our results indicate that monotonic assumptions in Earth system models may overlook important nonlinearities and potentially bias estimates of terrestrial carbon flux responses to seasonal climate extremes. Interpretable machine learning therefore provides a transparent framework for evaluating and improving the representation of soil carbon dynamics under climate change.