Application of EPR-MOGA in computing the liquefaction-induced settlement of a building subjected to seismic shake

Abstract

Accurate prediction of the liquefaction-induced settlement ( $${S}_{\mathrm{lc}}$$ ) is an essential requirement for a good design of buildings resting on liquefiable ground and subjected to seismic shake. However, prediction of the $${S}_{\mathrm{lc}}$$ is not straightforward process and it requires advanced soil models and calibrated soil parameters that are not readily available for designers/practitioners. In addition, the available empirical models to estimate the $${S}_{\mathrm{lc}}$$ have been developed using either classical regression analysis or multivariate adaptive regression splines and such techniques produce complicated models. Also, these empirical models have been developed utilizing results of numerical modelling. To overcome these limitations, novel model has been developed in this paper utilizing robust regression analysis driven by artificial intelligence called the evolutionary polynomial regression analysis. The new model has been developed using centrifuge results (real laboratory measurements) and can be easily used to accurately estimate the liquefaction induced settlement. The developed model scored a mean absolute error, root mean square error, mean, standard deviation of the predicted to measured values, coefficient of determination, $$a20 - \mathrm{index}$$ , and EPR coefficient of determination of 2.12 cm, 2.84 cm, 1.06, 0.19, 0.98, 0.77, and 97%, respectively, for the learning data and 1.73 cm, 3.31 cm, 0.99, 0.17, 0.97, 0.75, and 97%, respectively, for the examination data. The developed model has also been used in a parametric study to provide an insight into the sensitivity of the $${S}_{\mathrm{lc}}$$ to the foundation width, building height, pressure applied on the foundation, thickness and relative density of the liquefiable layer, and earthquake intensity. The results obtained from the parametric study are reasonable and in agreement with previous studies in the literature. Thus, the developed model can be employed to optimize designs and to reduce design costs as it does not require complicated analyses and/or expensive computational facilities.