Machine Learning Integrated with a Causal Pathway Framework Unravels Differential Mechanisms of Biochar-Driven Soil Organic Carbon Dynamics under Cadmium Stress

Biochar-mediated soil organic carbon (SOC) dynamics in cadmium (Cd)-contaminated soils are governed by complex interactions among biochar properties, soil characteristics, and environmental factors. However, the key drivers and causal mechanisms remain unclear, hindering the design of biochar strategies for carbon sequestration. This study integrated machine learning (ML) and partial least-squares path modeling (PLS–PM) to establish an interpretable causal framework. Using a global data set, a high-precision random forest model quantified the primary drivers. Soil properties dominated the predictions (60.27%), with phosphorus (P) (optimal level: <0.7 g/kg) and pH emerging as the most critical factors. Nonlinear thresholds showed that the Cd role shifted from positive to insignificant beyond 5.8 mg/kg. PLS–PM quantified the causal pathways: (i) physicochemical interactions (e.g., P competitive adsorption reduced SOC by β = −0.62, p < 0.001); (ii) climate-mediated cascades; and (iii) biochar aging feedback. Validation of the model using independent data sets beyond the training scope demonstrated a reasonable relative ΔSOC prediction error range (±6%–20%). SOC accumulation increased with higher pH, Cd content, and biochar dosage. Site-specific design should prioritize (i) inherent soil properties (pH and P), (ii) Cd gradients, and (iii) tailored biochar parameters. The framework enhances biochar design to optimize SOC sequestration and synergistic Cd management.