Explainable Machine Learning-Driven Targeted Modification of Biochar and Its Remediation Performance for Saline-Alkaline Soil.

Biochar is a low-cost, eco-friendly candidate for saline-alkali soil remediation. Its practical efficacy, however, is constrained by high alkalinity, an underdeveloped pore structure, and limited surface functional sites, all of which conventional modification methods have yet to adequately address. To bridge this gap, this study developed an explainable machine learning prediction system that integrates multi-source data spanning biochar preparation parameters, physicochemical properties, and remediation outcomes, thereby enabling the reverse design of modification processes for enhanced efficacy. From the SHAP feature-importance and dependence results, the recommended ranges for the key modification parameters were derived as follows: pyrolysis temperature=600-800 °C, particle size=10-70 mesh, specific surface area (SSA)=58-698 m2/g, electrical conductivity (EC)=349-3554 μS/cm, cation exchange capacity (CEC)=80-273 cmol/kg, pH=1.76-7.8, O/C ratio=0.36-0.56, and an addition ratio =2%-10%. The feature importance ranking was: EC > pH > SSA > CEC > particle size > O/C ratio > addition ratio > pyrolysis temperature. Based on the above results, this study proposed a phosphoric acid-chitosan composite modification process. The modified biochar exhibited an enriched pore structure (SSA = 21 m2/g), abundant oxygen-containing functional groups (O/C = 0.31), and low EC (625 μS/cm). Although its SSA and O/C ratio fell outside the recommended ranges, its EC was within the ideal range. Within 14 days, it reduced the saline-alkali soil pH from 10.26 to 9.28, whereas soil treated with unmodified biochar maintained a pH of approximately 10.2. This confirms EC as a core indicator and validates the reliability and practical applicability of machine learning-optimized preparation.