Performance Analysis of Boosting-Based Machine Learning Models for Predicting the Compressive Strength of Biochar-Cementitious Composites

Biochar, a carbon-rich material produced through the pyrolysis of wood residues and agricultural byproducts, has carbon storage capacity and potential as a low-carbon construction material. This study predicts the compressive strength of cementitious composites in which cement is partially replaced with biochar using machine learning models. A total of 716 data samples were analyzed, including 480 experimental measurements and 236 literature-derived values. Input variables included the water-to-cement ratio (W/C), biochar content, cement, sand, aggregate, silica fume, blast furnace slag, superplasticizer, and curing conditions. Predictive performance was evaluated using Multiple Linear Regression (MLR), Elastic Net Regression (ENR), Support Vector Regression (SVR), and Gradient Boosting Machine (GBM), with GBM showing the highest accuracy. Further optimization was conducted using XGBoost, Light Gradient-Boosting Machine (LightGBM), CatBoost, and NGBoost with GridSearchCV and Optuna. LightGBM achieved the best predictive performance (mean absolute error (MAE) = 3.3258, root mean squared error (RMSE) = 4.6673, mean absolute percentage error (MAPE) = 11.19%, and R2 = 0.8271). SHAP analysis identified the W/C and cement content as dominant predictors, with fresh water curing and blast furnace slag also exerting strong influence. These results support the potential of biochar as a partial cement replacement in low-carbon construction material.