AI-based approaches for predicting buried pipeline stability in cohesive-frictional soil under inclined forces: FELA, DNN, RNN, LSTM, and MARS

This study explores the uplift resistance of buried pipelines in cohesive-frictional soils using finite element limit analysis (FELA) combined with artificial intelligence (AI) techniques, specifically deep neural networks (DNN), recurrent neural networks (RNN), long short-term memory (LSTM) networks, and multivariate adaptive regression spline (MARS) for enhanced predictive modeling. Numerical simulations are conducted using the Mohr-Coulomb failure criterion as a soil strength material to assess the uplift capacity of pipelines, considering various parameters including embedment depth ratios ( H / D ), load inclination angles ( β ), adhesion factors ( α ), internal friction angles ( ϕ ), and soil strength ratios ( γH / c ). The findings demonstrate that increasing embedment depth and load inclination angle enhances uplift resistance by engaging larger volumes of surrounding soil. This paper also investigates and discusses the failure mechanisms of buried pipelines in cohesive-frictional soils, highlighting that horizontal loading ( β = 90°) induces the most significant soil mobilization. All proposed machine learning models were evaluated by rank analysis, performance parameters, and scatter plots, with the DNN model exhibiting higher performance ( R 2 = 99.98 %). Moreover, the sensitivity analysis identifies the embedment depth ratio as the most influential factor, followed by load inclination and adhesion. This thorough assessment illustrates the accuracy and reliability of the DNN model as an effective tool for predicting the uplift resistance of buried pipelines, delivering valuable insights for geotechnical applications in offshore and onshore settings.