Comparison Between Machine Learning Predictions and Eurocode 8 Empirical Expressions for the Plastic Rotation Capacity of Rectangular Rc Columns

Abstract

This paper presents a comparative assessment of the plastic rotation capacity of rectangular reinforced concrete columns, using machine learning predictions and the empirical expressions provided by Eurocode 8, Part 3. The study relies on a database of 258 rectangular RC column specimens subjected to quasi-static cyclic lateral loading, covering a wide range of reinforcement configurations, axial load levels, shear span-to-depth ratios and failure modes. A previously developed and validated Stacking Ensemble model, combining Support Vector Regression, Random Forest and XGBoost through a RidgeCV meta-regressor, is used as the data-driven reference model. Its predictions are compared with the values obtained using expressions A.3 and A.4 from EN 1998-3:2005, evaluated on a subset of 52 rectangular specimens for which all parameters required by the code expressions were available. The comparison is performed in terms of the coefficient of determination, mean absolute error and root mean squared error, together with a residual error analysis. The results show that the Stacking Ensemble model provides a closer agreement with the experimental values, while the code-based expressions exhibit larger dispersion and a more conservative response, particularly when the partial safety factor is considered. The findings indicate that data-driven models can complement empirical code provisions by providing an additional quantitative reference for the seismic assessment of existing rectangular RC columns.