Ensemble-learning surrogate models for the NORSOK M-506 CO₂ corrosion-rate model: Accuracy, stability, and SHAP-based interpretation

Abstract

Internal CO₂ corrosion is a major integrity concern in carbon-steel oil and gas pipelines, and models such as NORSOK M-506 are widely used to estimate corrosion rates under water-wet CO₂ service, though repeated evaluation is cumbersome for screening, uncertainty analysis, and optimization. This study develops surrogate models that approximate the NORSOK M-506 response surface using a public Monte Carlo-generated dataset of 243 operating cases, with temperature, internal pressure, CO₂ partial pressure, flow velocity, inhibitor efficiency, shear stress, and pH as inputs and radial corrosion rate as the target. Ridge regression served as a linear baseline, while Random Forest (RF) and XGBoost were evaluated as nonlinear ensemble models. Both ensembles reproduced the NORSOK response with very high accuracy, XGBoost performing best (holdout MAE = 0.0446 mm/y, RMSE = 0.0580 mm/y, R² = 0.9973; cross-validated R² = 0.9980 ± 0.001), closely followed by RF, whereas Ridge performed substantially worse and showed systematic residual curvature, indicating nonlinear interactions a linear model cannot capture. SHAP analysis identified inhibitor efficiency, temperature, and CO₂ partial pressure as the most influential predictors, followed by pH, with flow velocity, internal pressure, and shear stress contributing little within the sampled range. Ensemble learning thus provides accurate, stable, and interpretable surrogates for rapid approximation of NORSOK-based CO₂ corrosion predictions, within the sampled operating space and the assumptions of the underlying model. The dataset and surrogate describe uniform internal CO₂ corrosion of carbon steel in water-wet (sweet) service and do not apply to sour, H₂S-containing systems.