Probabilistic assessment of the seismic performance of earth slopes using computational simulation
Earthquake-induced permanent slope displacement has been the main damage measure used in evaluating the seismic performance of earth slopes and various predictive models for this displacement have been proposed. However, these predictive models are mostly based on displacements computed using sliding block analysis, although nonlinear finite element or finite difference simulations are becoming the preferred method to evaluate the performance of slopes. This research aims at developing predictive models for slope displacement based on nonlinear finite element simulations, and demonstrating how these predictive models can be used in probabilistic assessments of slope displacement. These methodological developments are demonstrated first using a single slope geometry representative of a site-specific analysis and then generic predictive models are established using a range of slope geometries. These generic displacement models are developed through both classical and artificial neural network (ANN) regression. Toward these goals, this research comprises the following three sections.
Nonlinear finite element analyses are performed for a soft clay slope using a suite of 105 input motions and the computed displacements are used to develop slope-specific displacement prediction models that utilize different ground motion intensity measures. The efficiency and proficiency of the displacement models using different combinations of intensity measures are assessed. These displacement models are used to compute probabilistic hazard curves of the permanent displacement, which represent the annual frequency of exceedance for a range of displacement levels. The computed hazard curves provide insight into the range of epistemic uncertainty associated with different displacement models.
A large set of nonlinear finite element simulations are performed on 40 slope models each subjected to more than 1000 input motions. A generic predictive model for displacement is derived from the computed displacements using classical regression techniques. The predictive model characterizes the slope in terms of its yield acceleration (ky) and the natural period of the sliding mass (Ts), and characterizes the input motion in terms of its peak ground velocity (PGV). The displacement variability is partitioned into the between-slope component, which represents the variability associated different slope models, and the within-slope component, which represents the variability due to different input ground motions.
Lastly, the database of slope displacements used in the classical regression are used to develop an artificial neural network (ANN) predictive model for displacement. ANN models allow researchers to investigate complex interactions between independent and dependent variables without specifying any restrictions on the functional form. The developed ANN moderately improves the displacement prediction relative to the classical regression model, although without the need of a complex functional form. The ANN displacement model is presented as a simplified mathematical expression that can be easily implemented into deterministic or probabilistic assessments of slope performance.