Compressive Strength Prediction of Basalt Fiber Reinforced Concrete Based on Interpretive Machine Learning Using SHAP Analysis

Abstract

Compressive strength prediction of Basalt Fiber Reinforced Concrete (BFRC), an advanced building material that combines performance and sustainability, is a complex task influenced by many factors. In this study, the compressive strength of BFRC is predicted using four tuned machine learning models, namely, Support Vector Machine (SVR), Random Forest (RF), Back Propagation Neural Network (BPNN), and Extreme Gradient Boosting (XGB), and analyzed using SHAP (Shapley additive approach). To build the machine learning model, a database containing 309 sets of BFRC compressive strength data collected from published articles was established in this study, and an additional 8 sets of BFRC compressive strength data were obtained through experimental work. SHAP interaction plots were generated to explain how the value of each characteristic affects the model prediction, and the optimal range of values for the basalt fiber characteristics was clarified. The results show that the XGB model outperforms the other three models in terms of prediction, with the coefficient of determination (R²) value of 0.9431, the root mean square error (RMSE) of 3.2325, and the mean absolute error (MAE) of 2.3355. Among the three basalt fiber parameters, the volume content of the basalt fibers has the greatest effect on the model output. In addition, the optimal range of volume content was 0.1%, the optimal range of diameter was 15–20 μm, and the optimal range of length was 8–15 mm.