Abstract: Characterizing the spatial variability of soil parameters with depth using stationary random fields is challenging, and three-dimensional (3D) slope reliability analysis is often time-consuming. Therefore, the stepwise Karhunen-Loève series decomposition method is employed to generate the stationary random field and the nonstationary random field considering the depth effect (the mean depth of undrained shear strength S_u is different when the random field is realized). A surrogate model based on sliced inverse regression (SIR) and extreme gradient boosting (XGBoost) is proposed, combined with Monte Carlo simulation (MCS) for 3D slope reliability analysis. Take a typical 3D slope, the effectiveness of the proposed method is verified. The calculation results based on stationary and nonstationary random fields are compared, and the influence of the burial depth parameter zb and the trend component b on the 3D slope reliability analysis results is discussed. The results indicate that the surrogate model proposed in this study can accurately and efficiently calculate the 3D slope failure probability. The burial depth parameter zb and trend component b significantly influence the slope reliability analysis results, and these parameter values should be selected reasonably according to the in-situ data. When the slope reliability is analyzed using nonstationary random field theory considering the depth effect, the failure probability increases as the trend component b increases. Conversely, without considering the depth effect, the nonstationary random field yields opposite results. The research results can provide an effective approach for the efficient reliability analysis of actual 3D slope.

Key words: three-dimensional slope, reliability analysis, spatial variability, extreme gradient boosting, nonstationary random field