ABSTRACT Biochar soil amendment is widely recognised as a promising strategy to improve soil quality and boost crop productivity; yet its field performance is highly variable—ranging from negative to positive effects—depending on biochar properties, soil type and agro‐climatic conditions, which restricts its broad‐scale adoption. To address this, we compiled 276 biochar amendment observations from nine countries spanning tropical, subtropical and temperate climates and developed nine Bayesian‐optimised machine learning models to predict relative changes in soil pH, soil organic carbon (SOC), cation exchange capacity (CEC) and crop yield. Results showed that among the nine models evaluated, CatBoost and XGBoost ranked as the top performers with distinct strengths, as CatBoost excelled in predicting CEC ( R 2 = 0.65) and pH ( R 2 = 0.86), whereas XGBoost was superior for SOC ( R 2 = 0.85) and crop yield ( R 2 = 0.66). Leveraging their complementary strengths, we developed an ensemble‐based visual prediction tool through collaborative prediction and real‐time intelligent allocation. In randomised validation tests, the tool achieved high precision, yielding mean relative errors of only 9.50% (SOC), 4.55% (CEC), 1.51% (pH) and 3.56% (crop yield). Feature importance and SHAP analysis revealed distinct regulatory modes—precipitation overwhelmingly dominated SOC responses, biochar application rate and duration served as primary levers for soil pH, while CEC and crop yield were governed by the integrated effects of multiple interacting factors. Partial dependence plots (PDP) further identified an optimal practice, showing that low‐temperature biochar (< 500°C) applied at ~20 t ha −1 every 2–3 years maximises soil‐crop benefits. Overall, by integrating local climatic conditions with soil characteristics, this framework provides a robust decision‐support system for targeted biochar soil amendment across diverse regions.
Xiang et al. (Wed,) studied this question.