Shap-interpretable prediction of nutrient leaching risks in biochar-based slow-release fertilizers

Biochar-based slow-release fertilizers (BSRFs) hold promise for improving soil nutrient use efficiency by reducing element release rates, making the assessment of their slow-release performance crucial. To optimize the assessment method and identify key driving factors, overcoming the limitations of traditional approaches such as high resource consumption and poor result comparability, a machine learning (ML) model for predicting nutrient leaching rates of BSRFs was developed based on existing literature data. Tree-based models including Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), Gradient Boosting Regression Tree (GBRT), and Light Gradient Boosting Machine (LightGBM) were employed with hyperparameter optimization through Tree-structured Parzen Estimator (TPE) algorithm, complemented by Individual Conditional Expectation (ICE) and SHapley Additive exPlanation (SHAP) for model interpretation. To evaluate the performance of the proposed models, two popular statistical indexes namely root mean square error (RMSE) and determination coefficient (R²) were used. Results show that the optimized LightGBM model achieved superior performance (RMSE = 2.2889, R² = 0.9946), with learning rate (0.63) and min_child_samples (0.30) identified as crucial hyperparameters. Nutrient leaching rate prediction is primarily driven by BSRFs element content (E-M) and leaching volume (V), with low model heterogeneity indicating prediction stability. High elemental content synergistically increased leaching risks with elevated leaching volumes, while enhanced specific surface area promoted nutrient retention. Offline and online graphical user interfaces (GUI) were developed to facilitate practical model implementation. The research provides data-driven decision support for BSRFs design and soil nutrient management.