Prediction of inelastic displacement demand for SDOF structures built on firm soils under near- and far-fault records using neural networks

The estimation of peak inelastic response demand for a single-degree-of-freedom structure (SDOF) through utilizing the corresponding SDOF peak elastic displacement is investigated. A comprehensive statistical analysis study of inelastic displacement ratios is presented for eighty ground motion records categorized into four ensembles representing National Earthquake Hazards Reduction Program site classes B, C, and D; near-fault records recorded on-site class D were also included. The parametric study is conducted for three hysteretic models: elastic-perfectly plastic, large-Takeda, and thin-Takeda models. The study showed that site-class has an insignificant effect on the inelastic displacement demand. However, near-fault records showed a significant increase in the inelastic demand, especially in the acceleration-sensitive region of the elastic response spectra. The hysteretic model has an insignificant effect on the inelastic demand for a period of more than 1 s; however, it showed a significant difference for shorter periods. Estimating the peak response of nonlinear structures constitutes a key aspect of seismic design; therefore, the development of accurate prediction models will reduce computational demand and makes the seismic analysis feasible practice. This paper deployed machine learning to predict inelastic displacement ratios through feedforward artificial neural networks. The inelastic demands obtained from the statistical analysis were used to train, validate, and test the neural network. Input parameters include the structure period and displacement ductility demand. The proposed architecture of the neural network showed significant accuracy in predicting inelastic responses. The proposed prediction model needed a smaller data sample for training as compared to traditional formulas obtained through nonlinear regression of large sets of data.