Machine Learning-Assisted Seismic Response Assessment of Regular and Irregular Buildings Under Long-Period Ground Motion

The ability to accurately predict seismic structural response is still a big challenge especially for regular and irregular buildings with long period ground motions. The time-history analysis of the system is nonlinear and requires a significant amount computation effort, hence this has motivated the use of Machine Learning (ML) based predictive methods. This study is an ML-aided methodology to assess maximum interstorey drift and roof displacement based on the structural properties and ground-motion parameters. The methodology involves numerical structural analysis, dataset preparation, feature engineering and regression-based prediction of numerical models namely Random Forest (RF), Extreme Gradient Boosting (XGBoost), Artificial Neural Networks (ANN) and Support Vector Regression (SVR). Five-fold cross validation and feature elimination analysis is used to investigate model performance. From the results it could be observed that fundamental period and spectral acceleration have the greatest effect on roof displacement. The importances of these features are 0.50 and 0.37 respectively for RF and 0.265 and 0.240 respectively for XGBoost. The feature elimination finds minimum RMSE values of about 5.1 to predict roof displacement with the inclusion of three predictors and 0.047 to predict maximum interstorey drift with four predictors. The results here show that feature selection could enhance the simplicity of the model and its prediction performance. The proposed framework is a promising complementary method for a quick seismic response assessment but needs to be further validated against independent datasets before it could be implemented in practice.