Seismic response prediction of irregular buildings using machine learning: a comparative analysis of parametric and non-parametric models

Abstract

The increasing demand for aesthetic designs, coupled with limited land availability, has resulted in irregular building configurations that compromise seismic performance. These irregularities can lead to stress concentrations caused by torsional effects and variations in stiffness. This study aims to predict key seismic responses such as natural time period, displacement, and storey drift for geometrically irregular reinforced concrete (RC) buildings. A total of 630 building models were developed and analyzed using ETABS, incorporating input parameters like the structural coefficient (r), building height (H), and irregularity index (β). We applied various machine learning (ML) algorithms, including parametric models such as Multiple Linear Regression, Ridge, and Bayesian Ridge, as well as non-parametric models like Decision Tree, Random Forest, AdaBoost, XGBoost, and Gaussian Regressor. Model performance was evaluated using metrics such as R², Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and K-fold cross-validation to ensure robustness. Among the models assessed, Gaussian Boosting demonstrated superior performance with an R² value of 0.99347, while Ridge Regression exhibited the lowest accuracy. This study highlights the effectiveness of machine learning techniques, particularly Gaussian and multi-linear models, as accurate, fast, and cost-effective alternatives to traditional methods of predicting seismic response. In addition, we introduced a graphical user interface (GUI) as a user-friendly tool designed to assist researchers in estimating the seismic capacity of reinforced concrete buildings. This GUI aims to minimize computational demands and reduce the complexity of analytical procedures.