Leveraging AI and Physics‑Based Modeling to Predict Internal Corrosion Hotspots in Crude‑Oil Export Pipelines (C2026-00326)

Sridhar Arumugam, Patrick Teevens, Sureshkumar Srinivasan, Bharanidharan Hemachandran

Traditional internal corrosion assessments in liquid pipelines rely on combining flow modeling with inspection data to identify root causes and corrosion mechanisms. However, unexpected failures are often driven by localized corrosion in discrete areas where water or solids accumulate, creating aggressive microenvironments. These localized forms of attack are complex and often fall outside the predictive scope of current standards.
UT-ILI detected extensive metal loss in two pipeline sections transporting export-quality crude oil. This study presents a hybrid workflow that combines Broadsword’s in-house physics-based internal corrosion prediction model, enpICDA™, with supervised machine learning algorithms to quantify relationships between operating conditions and observed corrosion damage.
The physics-based model was first executed under representative scenarios to simulate axial pressure and temperature profiles, in-situ velocities, inclination angles, and the locations of water holdup and solids deposition. These outputs were then integrated with UT-ILI metal loss data, resulting in over 2,900 labeled data points.
Gradient-boosted trees, random forests, and linear regression were trained and cross-validated to rank the influence of each parameter on internal corrosion rates. Gradient-boosted trees achieved higher R² values. This integration of physics-based modeling with machine learning offers a promising framework for proactive and cost-effective corrosion management in critical pipeline infrastructure.