Prediction of ground movement-induced pipe responses considering variable PGD magnitudes using physics-informed neural networks and transfer learning

Long-distance buried pipelines are susceptible to Permanent Ground Displacements (PGDs) triggered by geo-hazards. Although pipeline routing is generally considered aiming to avoid areas with challenging geological conditions, this is not always feasible or possible due to economic, technical, and environmental constraints and uncertainties encountered by decision-makers. Therefore, the structural response of buried pipelines under PGDs is a major concern regarding pipe’s safety and integrity in the industry. This paper introduces a novel approach within a deep learning framework, specifically using a Physics-Informed Neural Network (PINN), to (1) predict the elastic response of buried pipelines, such as Carbon Fibre Reinforced Polymer (CFRP) pipelines or steel pipelines in the elastic stage, under permanent ground displacement of varying magnitudes using a unified model and (2) evaluate the effects of different parameters, such as ground movement magnitude and direction, on pipeline responses. The accuracy of the predicted results is verified against two conventional numerical approaches, namely Finite Element (FE) and Finite Difference (FD) methods. The results and findings of this research demonstrate the promising capability of the PINN method in predicting both displacement and strain fields of pipelines subjected to a range of PGD magnitudes. The proposed approach using PINN can serve as a meshless, cost-effective, and simulation-free alternative for pipeline response prediction in integrity assessment. • A PINN model integrated with transfer learning predicts the response of buried pipelines under varying ground movement. • The PINN framework provides a simulation-free, mesh-free, and cost-effective approach for pipeline integrity assessment. • The study considers the effects of geometric nonlinearity and nonlinear pipe-soil interaction on pipeline responses. • Transfer learning within the PINN model significantly reduces training time. • Results show that ground movement magnitude and crossing angle substantially affect strain and deformation responses.