Plant carbon fluxes govern soil organic carbon dynamics under climate change: Machine learning reveals critical GPP thresholds

Soil Organic Carbon (SOC) dynamics are profoundly influenced by climate change, yet the contributions of plant carbon fluxes, climatic forcing, and land-use changes remain poorly quantified. Here, we develop a machine learning framework to assess the drivers of SOC dynamics at both 0-20 and 20-100 cm depths across China from 1980 to 2100. The proposed model, validated via 10-fold cross-validation, achieved R 2 values of 0.41 and 0.48 for the 100-cm and 20-cm soil profiles, respectively. Furthermore, independent temporal validation across two distinct periods yielded acceptable performance, demonstrating the model’s robust capacity to capture both the spatial distribution and temporal dynamics of SOC stocks. Future projections under four Shared Socioeconomic Pathway (SSP) scenarios indicate a national-scale SOC increase of 4.8 to 7.7 Pg C at 0-100 cm by 2100 compared to 2000-2015. Sensitivity analyses show that plant carbon fluxes are the dominant control on SOC variability. SHapley Additive exPlanations (SHAP) analysis further reveals a nonlinear Gross Primary Productivity (GPP)-SOC relationship that partially offsets warming-induced SOC losses. This relationship is characterized by two key thresholds. First, at GPP ≈ 4 gC m -2 d -1 , the marginal contribution of GPP to SOC (SHAP value) transitions from negative to positive, indicating the ecosystem transitions from a net carbon deficit to a surplus, coinciding with the ecologically sensitive 400 mm isohyet of equivalent precipitation. Below this threshold, insufficient inputs stimulate existing SOC decomposition (likely priming effect). This zero-crossing indicates a regime shift where the GPP-driven carbon input overcomes historical deficit. Second, at GPP ≈ 7 gC m -2 d -1 , the positive GPP-SOC sensitivity plateaus, indicating diminishing returns in carbon sequestration with further increases in productivity. Under high-emission scenarios, this transition zone expands northwestward into historically arid regions. These findings provide actionable insights for climate-resilient land management, highlighting targeted vegetation restoration, especially afforestation aligned with hydroclimatic thresholds, as a nature-based solution to safeguard SOC stocks in a warming world. • Under climate change, carbon fluxes have a substantial and non-negligible impact on SOC dynamics. • A nonlinear GPP-SOC relationship features a threshold (GPP≈4 gC m -2 d -1 ), switching from carbon-input deficit to surplus. • Under high carbon emission scenarios, the GPP-SOC transition zone expands northwestward into arid regions in China.