Predictive equations for earthquake-induced dynamic response and permanent displacements of slopes considering bedrock depth and impedance contrast

The bedrock depth and soil-bedrock impedance contrast are usually neglected in existing predictions of earthquake-induced slope displacement (D). This study develops predictive models for equivalent seismic loading parameters (that measure the dynamic response of sliding mass) and D, based on over 267,000 dynamic response analyses and more than 7 million sliding block analyses. The models are advantageous over the existing ones in: (i) incorporating the distance between slip surface and bedrock along with the effects of dynamic response and soil-bedrock impedance contrast; (ii) introducing artificial neural network (ANN) to improve predictive accuracy compared to classical functional forms; (iii) utilizing pseudo-spectral accelerations (SAs) at specific periods as alternative predictors to the less readily available mean period (T_m), with a systematic comparison between the T_m- and SA-dependent predictions. Multiple parallel models, incorporating various intensity measures (IMs) as predictor variables, are provided to potentially account for epistemic uncertainty and address situations where seismic hazard information is limited to specific IMs, while the newly introduced IM-vector [peak ground acceleration, spectrum intensity] is recommended. The models generally yield unbiased predictions without overfitting. Moreover, classical functional forms are utilized to construct reference models employing the same predictors and data for comparative purposes. The results reveal that ANN decreases the standard deviations of loading parameters and D by 25–60 % and 5–20 %, respectively. This article not only provides explorative insights into the machine learning application, but also offers a practical tool for the quick evaluation of seismic slope performance on both site-specific and regional scales.