An efficient computational framework for seismic reliability analysis of submarine slopes

Assessing the seismic stability of submarine slopes remains challenging due to the inherent spatial variability of soil properties and the stochastic nature of earthquake-induced loads. While Monte Carlo Simulation (MCS) is robust for probabilistic analysis, it becomes computationally expensive when combined with complex random field discretizations. The Karhunen-Loève (KL) expansion, widely used to model spatial variability, often results in high-dimensional input features, leading to the “curse of dimensionality” for traditional surrogate models. To overcome these limitations, this paper proposes a hybrid surrogate modeling framework combining Partial Least Squares (PLS) for dimensionality reduction and Support Vector Regression (SVR) for nonlinear prediction. A high-dimensional dataset (203 input features) considering both spatially variable shear strength parameters and seismic characteristics was generated based on the pseudo-dynamic Bishop’s method. The analysis demonstrated that PLS effectively compressed the 203 input features into just 9 latent variables. The resulting KL-PLS-SVR model achieved an R2 of 0.9751 and an RMSE of 0.0300, reducing the prediction error by 19.8% compared to KL-PLS linear regression. This framework offers a computationally efficient and accurate tool for seismic reliability analysis of submarine slopes involving high-dimensional spatial variability, with potential applications in offshore and coastal engineering practice.