Prediction of the fundamental period of infilled reinforced using advanced machine learning methods

Abstract

The use of machine learning (ML) to solve civil engineering problems has increased remarkably during the last few decades due to its effectiveness in reliably approximating complex relationships. In this paper, a key parameter of seismic design is estimated using hyperparameter ML algorithms to develop predictive models that compute the fundamental period. Initially, the impact of the train-test split ratio was investigated using three different splits, where the best results were achieved with train-test split ratios equal to 90/10 for all metrics. By predicting the fundamental period with three ML methods, namely XGBoost-HYT-CV, DANN-MPIH-HYT, and RF-HYT, the best fit was acquired by XGBoost-HYT-CV (coefficient of determination R2 = 99.994% and mean absolute error MAE = 0.00428). Although international literature agrees that building height is the primary factor influencing the fundamental period, feature engineering has revealed that the natural logarithm of the percentage of openings is the most significant parameter. This finding underscores the value of feature engineering in generating additional variables and uncovering their impact on output variables. Finally, an equation was derived from POLYREG-HYT that outperformed all existing formulae, deriving a final MAE of 0.0153, approximately three times smaller than the best-performing equations proposed in the international literature.