ANN and XFEM-Driven Predictive Framework for Axial Crack Failure in Pipelines: Advancing beyond CorLAS

The CorLAS model is widely used in the pipeline industry for predicting the failure pressure of pipelines with axial surface cracks, primarily due to its low computational and moderate prediction accuracy compared to other alternative models. Despite its popularity, recent studies have reported that the CorLAS model is associated with a high coefficient of variation (CoV) of model error, which can undermine the confidence in reliability-based assessment of burst pressure of pipelines using CorLAS. Recently, machine learning techniques, such as artificial neural networks (ANN), have been successfully used to develop predictive models based on large datasets. Experimental burst testing involves significant costs and complex setups and is subject to strict regulatory constraints. As a result, publicly available datasets are limited, making developing ANN models based on experimental data infeasible. To address this issue, a sufficiently large database is generated using the extended finite-element method (XFEM), which has proven to be a powerful and effective approach to simulating crack growth. This study develops a hybrid ANN-XFEM predictive framework for failure pressure estimation in pipelines with axial cracks, improving upon existing empirical models such as CorLAS. A synthetic dataset of 498 XFEM simulations was generated, covering four pipeline grades (X52, X60, X65, and X80) with varying diameter-to-thickness ratios (D/t=40–90), crack depths (a/t=0.4–0.8), and crack lengths (L2/Dt=1–18). The trained ANN model achieved a mean absolute error of 0.43 and a CoV of 4.76%, significantly outperforming CorLAS (CoV=14%). These findings offer a computationally efficient alternative to full-scale burst testing and high-fidelity XFEM simulations, enabling real-time integrity assessment and predictive maintenance for pipeline operators.