Machine learning techniques to predict the fundamental period of infilled reinforced concrete frame buildings

Predicting the structure's fundamental period is a challenging task since its value changes when the features of buildings change. However, it is more cumbersome for reinforced concrete buildings with infill walls due to their nonlinear behavior. The objective of this study is to highlight the importance of hyperparameter optimization in machine learning models to achieve the best performance in predicting the fundamental period of infilled RC frame buildings. To this end, three types of boosting algorithms have been chosen, namely: gradient boosting decision trees, lightGBM, and catboost. Compared to the default parameters for each model, the fine-tuning hyperparameter models with the Optuna framework yielded the best results, with a higher coefficient of determination and lower error values. Furthermore, the lightGBM model set by its optimized parameters out-performed the other two boosting methods. The feature selection technique emphasizes the five parameters' contributions in estimating the building's fundamental period by demonstrating the effect of removing one or more parameters each time on determining this dynamic characteristic. Moreover, the multivariate adaptive regression splines is used to develop a mathematical expression for the fundamental period that outperformed building code formulas and some authors' proposed equations. Finally, a set of smooth curves derived from the optimum lightGBM and MARS models were introduced, revealing that the models were not overfitting.