Artificial neural network-aided decoupled prediction of earthquake-induced shallow and deep sliding displacements of slopes

Abstract

The Newmark-type predictive models are extensively used to estimate earthquake-induced sliding displacements (D) of slopes. Most existing models are designed for shallow slope failure using polynomial regression. Based on a large amount of decoupled sliding-block analyses, this paper proposes an artificial neural network (ANN)-aided model to predict D for both shallow and deep slope failures using peak ground acceleration and spectral acceleration at the 2 s period (SA(2 s)). Three sub-models are included for estimating shallow sliding displacement, representing dynamic response of sliding mass, and modifying the displacement for deep failure, respectively. The key features achieved are as follows: (1) the inputted SA is more easily accessible than the mean period (Tm) required by the existing models; (2) impedance ratio (IR) is utilized to account for stiffness conditions of geo-materials underlying slip surfaces; (3) powerful ANN is introduced for the first time as a surrogate of the decoupled analysis. The SA-based model generally yields lower biases and uncertainty than the existing models. Since larger D is produced for smaller IR, the existing models with a single IR value would be unconservative for relatively stiff underlying soils and/or soft overlying soils. Coefficients of the proposed model are provided for practical applications.