Machine Learning-based prediction of seismic lateral deflection of steel trusses using nonlinear time-history analysis

This study evaluates the effectiveness of nine machine learning (ML) algorithms, including Gaussian Process (GP), support vector machines (SVM), backpropagation feed-forward neural network (BP-FNN), random forest (RF), gradient tree boosting (GTB), adaptive boosting (AdaBoost), Extreme gradient boosting (XGBoost), Light gradient boosting machines (LightGBM), and categorical boosting (CatBoost), in predicting seismic lateral deflection of truss structures using nonlinear time-history analysis. Four numerical examples involving different truss types subject to the El Centro earthquake are examined. Key findings show that seismic lateral deflection of truss structures has a non-normal distribution with heavy tails. Tree-based ensemble algorithms, particularly CatBoost, provide more accurate predictions than GP, SVM, and BP-FNN. The complexity of such prediction problems may require larger training samples; however, computationally intensive algorithms like GP, SVM, GTB, AdaBoost, and BP-FNN may be less effective. LightGBM is notable for maintaining efficiency even with larger datasets. CatBoost demonstrated the highest reliability and accuracy, followed closely by LightGBM. Future research will address additional aspects like the impact of variable predefined design spaces and applied loads as design variables and extend the current work to structural reliability analysis and optimization problems.