Stochastic seismic slope stability assessment using polynomial chaos expansions combined with relevance vector machine

This paper presents probabilistic assessment of seismically-induced slope displacements considering uncertainties of seismic ground motions and soil properties. A stochastic ground motion model representing both the temporal and spectral non-stationarity of earthquake shakings and a three-dimensional rotational failure mechanism are integrated to assess Newmark-type slope displacements. A new probabilistic approach that incorporates machine learning in metamodeling technique is proposed, by combining relevance vector machine with polynomial chaos expansions (RVM-PCE). Compared with other PCE methods, the proposed RVM-PCE is shown to be more effective in estimating failure probabilities. The sensitivity and relative influence of each random input parameter to the slope displacements are discussed. Finally, the fragility curves for slope displacements are established for sitespecific soil conditions and earthquake hazard levels. The results indicate that the slope displacement is more sensitive to the intensities and strong shaking durations of seismic ground motions than the frequency contents, and a critical Arias intensity that leads to the maximum annual failure probabilities can be identified by the proposed approach.