Existing code equations for predicting the bending capacity of ultra high performance fiber reinforced concrete (UHPFRC) beams often show large scatter, with coefficients of variation (CoV) exceeding 30 to 60 %, leading to both overly conservative and unconservative estimates. This study develops a novel and explainable machine learning (ML) framework for accurate bending capacity prediction, representing the first systematic benchmarking of advanced ensemble ML methods against international and national design codes. An updated database of 264 experimental UHPFRC beam tests was compiled from the literature and partitioned into training (70%), validation (15%), and testing (15%) subsets. Six ensemble algorithms were optimized using Bayesian hyperparameter tuning with 10 fold cross validation, namely Random Forest (RF), Gradient Boosting Machine (GBM), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), and Extreme Gradient Boosting (XGBoost). The best performing models, CatBoost and XGBoost, achieved excellent predictive accuracy on unseen test data (R2 ≈ 0.96 to 0.97, RMSE ≈ 15 to 18 kN·m, CoV ≈ 10%) with essentially no systematic bias. These results clearly outperform code equations (R2 ≈ 0.60–0.70, CoV ≈ 28–64%, and biases up to 40 %). Interpretability analysis using Shapley Additive Explanations (SHAP) confirmed that effective depth and reinforcement ratio are the dominant predictors of bending strength, followed by steel yield strength and section properties, while UHPC compressive strength and fiber parameters had relatively minor influence within the dataset range. The novelty of this study lies in presenting an integrated and interpretable ML framework that not only achieves superior predictive performance but also provides mechanistic insight into UHPFRC beam behavior. The proposed approach offers a reliable data driven complement to current design codes and has potential for practical adoption in structural engineering design and code development.
Silewu et al. (2025) studied this question.