Flexural capacity prediction of partially encased composite beams using explainable machine learning

To address traditional limitations, this study investigated the flexural performance of large-section PEC beams with varied web openings using experiments and machine learning (ML). Four-point bending tests on specimens with different sections and openings demonstrated excellent ductility (coefficient >4.0), although openings slightly reduced yield strength without significantly affecting overall performance. A database of 15 variables was used to train and validate four ML models (RF, CatBoost, KNN, LightGBM). The RF model achieved the highest accuracy (~2.6% MAE). Shapley analysis improved interpretability by identifying key parameters. Integrating explainable ML substantially enhances the prediction accuracy and interpretability of PEC flexural capacity, offering a promising approach for intelligent structural design and assessent.