A novel hybrid surrogate model for the stability and post-failure analysis of spatially variable slopes using a smoothed sequential limit analysis

This study presents a novel framework for evaluating slope stability in spatially variable soils by integrating a newly developed sequential limit analysis based on the Hellinger-Reissner functional, utilizing the node-based smoothed finite element method (NS-FEM), with a newly proposed deep learning (DL) approach termed multi-downsampling hybrid Linformer-convolutional neural networks (CNNs). The NS-FEM-based mixed formulation of limit analysis (MFLA) enhances computational accuracy and convergence by smoothing strain fields and mitigating numerical discontinuities commonly encountered in standard finite element methods (FEMs). This method generates reliable datasets for stochastic simulations of slope stability under both static and seismic loading conditions. To address the computational expense of specific simulations, we propose the multi-downsampling hybrid Linformer-CNN model, a sophisticated DL architecture that employs dual parallel pathways with distinct downsampling strategies – AveragePpooling1D for medium-scale feature extraction and MaxPooling1D for coarse-scale feature extraction. Each pathway integrates one-dimensional (1D) CNNs for local feature extraction and Linformer-based self-attention mechanisms to efficiently capture global dependencies. The parallel downsampling strategies balance computational efficiency with feature granularity, enabling the model to leverage both local and global data characteristics effectively. The extracted multi-scale features are concatenated and further processed through fully connected networks (FCNs) to accurately predict the factor of safety (FoS) of slopes. Comparative analyses demonstrate that the hybrid Linformer-CNN model outperforms traditional FCN and CNN architectures, achieving robust and precise predictions with a mean absolute percentage error (MAPE) below 10%. Additionally, the proposed framework significantly reduces computational time, highlighting the potential of integrating NS-FEM-based MFLA with advanced DL architectures for rapid and reliable slope stability assessment in geotechnical engineering.