Interpretable Machine Learning for Biochar Ammonium Nitrogen Adsorption: Decoupling Synthesis and Operational Drivers and Quantifying Optimal Design Rules

Predicting the ammonium nitrogen (NH4+-N) adsorption capacity of biochar is challenged by coupled, nonlinear interactions among operational variables, pyrolysis synthesis parameters, and intrinsic material properties. A unified data set comprising 425 literature-reported adsorption records from 18 studies was compiled, and preprocessing (KNN imputation and Z-score standardization) was fitted exclusively on the training subset to prevent data leakage. Among seven machine learning algorithms benchmarked with Bayesian optimization and 5-fold cross-validation, XGBoost achieved the highest test-set coefficient of determination (R2) of 0.901 and root-mean-square error of 0.555 mg g–1. TreeSHAP analysis quantitatively decoupled three driver groups, adsorption operational parameters (53.3%), biochar physicochemical properties (38.2%), and pyrolysis synthesis parameters (8.5%), and identified initial concentration, biochar oxygen content, and adsorbent dosage as the dominant variables, with oxygen content exhibiting a unimodal relationship with adsorption capacity. Partial dependence plots were further utilized to reveal nonlinear response behaviors and determine data-coverage-validated optimal operation and design windows for key variables. A graphical user interface was developed to estimate equilibrium adsorption capacity, reducing reliance on extensive laboratory trials.