This study investigates the stability of three-dimensional (3D) earth slopes with complex topographies via machine learning models, specifically deep neural networks (DNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). A dataset comprising 500 samples was generated using the Scoops3D computer program, incorporating cohesion (c), internal angle of friction (ϕ), unit weight of soil ( \(\gamma\) ), and pseudostatic horizontal loading ( \({k}_{h}\) ) as input parameters and the factor of safety (FOS) as the output parameter. The analysis is based on the case history of the 1980 Mount St. Helens slope failure in Washington, USA. The critical surface failure is determined using the box search method, and the limit equilibrium method (Bishop's simplified method) is used to determine the factor of safety. The results of the sensitivity analysis performed on the dataset for the performing model indicate that the parameters c, ϕ, and \(\gamma\) have similar significances and that the \({k}_{h}\) effect has a relatively less significant parameter in estimating the FOS for the particular dataset. The novelty of this study lies in developing and applying advanced neural network models (DNNs, RNNs, CNNs) trained on 3D slope stability assessments for the first time at Mount St. Helens, which consider both seismic and nonseismic conditions. The predictive performance of the proposed models was analyzed via various performance indices, rank analysis, and an error matrix. An analysis of the obtained results reveals that the DNN model (R2 = 0.999 in training and R2 = 0.997 in testing) yields more accurate results than the CNN and RNN models do in predicting the safety factor of the 3D earth slope.
