Big data generation and comparative analysis of machine learning models in predicting the fundamental period of steel structures considering soil-structure interaction

Abstract

The computing of the fundamental period of structures during seismic design is well documented in design codes but is mainly dependent on the height of the structure, which is considered to be the most influential parameter. It is, however, important to consider a phenomenon called the soil–structure interaction (SSI), as this has been found to have a detrimental effect, especially for buildings founded on soft soils. A pilot research project foresaw the use of machine learning (ML) algorithms trained on relatively limited datasets for the development of a more accurate and objective fundamental period formula. Therefore, a dataset that consists of 98,308 fundamental period data points was created through the use of a High-Performance Computer (HPC), which is the largest dataset of its kind. The HPC results were then used to train, test, and validate different ML algorithms. It was found that XGBoost-HYT-CV with hyperparameter tuning performed the best with a correlation of 99.99% and a mean average percentage error (MAPE) of 0.5%. Furthermore, the XGBoost-HYT-CV model outperformed all under-study ML models when using an additional dataset that consisted of out-of-sample building geometries and soil properties, with a resulting MAPE of 9%. Finally, irregular buildings were also used to test the performance of the proposed predictive models.