The engineered cementitious composite (ECC)-reinforced concrete (RC) composite beam is an innovative structure, and its design and implementation rely on little experimental data. The machine learning (ML) technique, as an adjunct, can efficiently broaden the scope of data. This paper aims to present a novel approach for predicting the bending capacity of ECC-RC beam designs.
A dataset comprising 210 ECC-RC composite beam test results was utilized to train the model, employing ten techniques for prediction. The Shapley additive explanation (SHAP) method evaluates the influence of input features on prediction results by utilizing a high-performing ML model. Additionally, the numerical model is employed in this study to augment and further corroborate ML. The numerical model is employed to meticulously examine the significant input features of ML, so enhancing the comprehension of how input characteristics affect the bending capacity of composite beams.
The results underscore the effectiveness of ML prediction models in precisely estimating the bending strength of ECC-RC composite beams. The CatBoost model demonstrated superior prediction accuracy and generalization ability, with R2 values of 0.971 for the training sample and 0.978 for the test sample.
The SHAP analysis revealed that the cross-sectional area of the tensile reinforcement bar (), bottom ECC depth () and beam depth () were identified as the primary influencing features affecting the bending capacity of ECC-RC composite beams. Moreover, both the CatBoost and finite element (FE) model indicated that positioning the ECC layer near the base of the beam significantly enhances bending strength compared to its placement at the top.
