A Neural Network-Enhanced Finite Difference Approach for Strain Demand Prediction of Inelastic Pipes Subjected to Ground Displacement

Abstract Permanent ground displacement induced by geohazards poses a significant threat to the integrity of pipelines due to the potential for excessive strain. Accurately predicting strain demand is critical for guiding the design of new pipelines and assessing the risks associated with existing ones crossing geohazard zones. Previously, a finite difference approach for strain demand prediction in pipes subjected to permanent ground displacement was developed by the authors. It has been proven to be a simple and valuable technique for practical use in the pipeline industry, compared with conservative empirical formulas and time-consuming finite element modeling. However, the existing method relies on explicit expressions for axial force and bending moment, derived under the assumption of a bilinear stress-strain curve for the material, which restricts its applicability when dealing with more complex constitutive models that require numerical integration. To remedy this situation, a novel approach that models constitutive law using deep neural networks is proposed, serving as an alternative means for capturing stress-strain relationship. This novel approach is integrated into the finite difference scheme to overcome the constraints of the original method, thereby enhancing its applicability. A comprehensive case study was conducted to evaluate the effectiveness of the proposed neural network-enhanced finite difference approach in comparison to the original method. Results from this study demonstrated that the proposed method can achieve comparable accuracy to the original finite difference method when dealing with small ground displacements. The finding indicates the potential advantages of the proposed method in efficiently handling more complex constitutive relations, which will be explored in future work.