Regression Machine Learning Models for Probabilistic Stability Assessment of Buried Pipelines in Spatially Random Clays

The uplift capacity of pipelines buried in clay is a critical aspect of their structural integrity, affecting their stability and performance under varying conditions. This study investigates the probabilistic solutions of pipeline stability considering the spatial variability of soil strength by integrating random field theory and adaptive finite element limit analysis, known as random adaptive finite element limit analysis (RAFELA). In this paper, RAFELA is carried out to derive the statistical characterizations of pipelines under uplift force. Several advanced regression machine learning techniques, namely, emotional neural network (ENN), the group method of data handling (GMDH), adaptive neuro-fuzzy inference system (ANFIS), and minimax probability machine regression (MPMR), have been used to develop surrogate models to assess the mean uplift capacity factor of pipelines buried in spatially random clay subjected to both horizontal and vertical forces. The dimensionless parameters utilized in the present study include the embedment depth ratio (H/D), overburden factor (gamma H/mu c), adhesion factor (alpha), inclination angle (beta), coefficient of variation (CoVc), and dimensionless correlation length (Theta c), where Monte Carlo simulations are employed to perform the stochastic analysis. In addition, four advanced regression machine learning models are also utilized to develop surrogate models for estimating the mean uplift capacity factor of pipelines. Based on the performances of all the machine learning models, the proposed GMDH model is the most accurate model, with R2=0.9772 for training and R2=0.9370 for the testing phase, followed by the MPMR, ANFIS, and ENN models. The REC curve demonstrates the GMDH model's superiority, showing lower AOC values (training: 0.1715, testing: 0.2274) followed by MPMR (training: 0.2163, testing =0.2491), ANFIS (training:0.2408, testing: 0.3263) and ENN (training: 0.3177, testing: 0.3595) models.