Prediction of Biochar Phosphorus Slow-Release Rate Based on Ensemble Learning: Model Performance Comparison and Investigation of Controlling Factors for Slow-Release Performance

Aiming at the problem that the sustained-release performance of biochar is difficult to efficiently optimize through traditional experiments, this study constructs a biochar phosphorus slow-release rate prediction framework based on ensemble learning. Ten features such as pyrolysis temperature, retention time, microwave power, and elemental composition were selected as inputs, and the prediction performances of five integrated algorithms, namely, Random Forest, AdaBoost, XGBoost, GBDT, and LightGBM, were systematically compared. The results show that the GBDT model exhibits the best performance on the test set (R2 = 0.7395, MSE = 0.0071) and can effectively capture complex nonlinear relationships and suppress overfitting. The analysis of feature importance reveals that the pyrolysis temperature is the core factor influencing slow-release behavior, followed by hydrogen content, phosphorus content, and microwave power. Experimental verification on biochar slow-release fertilizers of apple wood, walnut shells and corn stalks demonstrated no significant discrepancy between the model predicted values and the measured values (p > 0.05). This study provides a new data-driven approach for the rapid prediction and mechanism analysis of the slow-release performance of biochar.