Physically constrained data augmentation and interpretable machine learning for predicting the flexural strength of basalt fiber-reinforced concrete

The flexural strength of basalt fiber-reinforced concrete (BFRC) is governed by nonlinear interactions among mixture proportions and fiber-related variables. To address the limited availability of BFRC data and the risk of overfitting, this study developed a workflow combining physically constrained generation–screening data augmentation, comparative machine learning, model interpretation, and graphical user interface (GUI) implementation. A literature-derived database of 245 real samples was divided into 172 training samples and a fixed real test set of 73 samples. Candidate generation and screening were confined to the training data, whereas the real test set was excluded from augmentation, synthetic-sample screening, hyperparameter configuration, and cross-validation. On the fixed real test split, Histogram-based gradient boosting (HGB) achieved the best numerical performance among the nine evaluated configurations, with a weighted training R² of 0.996 and an unweighted test R² of 0.931, together with test MAE and RMSE values of 0.261 and 0.423 MPa, respectively. SHAP identified silica fume, cement, and water as the leading predictive variables. One- and two-dimensional ALE analyses further showed that binder composition, water content, aggregate proportions, and fiber content jointly shaped the model response. The HGB-based GUI therefore provides preliminary BFRC mixture screening within the sampled data domain, while extrapolated or critical mixtures require experimental validation.