Prediction of Effective Stiffness in Rectangular RC Columns Using Ensemble and Boosted Machine Learning Models

Abstract

The Reinforced Concrete (RC) columns are one of the main structural elements that estimation of their effective stiffness is essential for trustworthy structural system design and analysis. As traditional approaches, both analytical and experimental, often involve simplifications that may not accurately represent real behavior, Machine Learning (ML) approaches offer a viable substitute. This work presents a flexible and high-performance ML-based ensemble framework in phyton environment based on the Stacking and Averaging techniques to improve the prediction of effective stiffness of RC structures. For developing ML models, this study utilizes a comprehensive dataset with 266 samples (80% for training and 20% for testing), which includes axial load capacity, the percentage of reinforcing steel, and the column dimensions and effective stiffness factors. The ML algorithms are trained through k-fold cross-validation and optimized using Grid Search-based hyperparameter tuning to improve prediction accuracy. The Stacking and Averaging ensemble models results the highest predictive accuracy in test-set with R2 values of 0.9708 and 0.9840, in comparison with Voting, Light Gradient Boosting, DecisionTree, and K-Nearest Neighbors models. Furthermore, the ensembles robustness in capturing nonlinear relationships among features are further confirmed by its consistently low test RMSEs of 0.0182 and 0.0169 values. Moreover, the interpretability analysis makes it clear that the percentage of steel reinforcement is the most influential parameter that have a significant contribution on RC’s effective stiffness. This finding demonstrates that such models have great potential for integration into structural design RC processes, enhancing efficiency and reliability in engineering practice.