Risk-Informed Corrosion Prediction for Supercritical CO2 Pipelines Using Nonlinear Machine Learning and Monte Carlo Uncertainty Propagation

Abstract

The integrity of Supercritical CO2 (SC-CO2) transport pipelines is threatened by the high temperature–pressure operating conditions, the presence of moisture, and aggressive impurities such as acids, alkalines, salts, O2, SO2, H2S, and NO2. For this reason, this environment gives rise to highly nonlinear and synergistic inconclusive corrosion mechanisms that cannot be captured by conventional CO2 corrosion prediction models. Widely used industrial models, including DeWaard95, BP Cassandra, Predict, and NORSOK, misrepresent the corrosion severity, creating a dangerous gap between predicted and actual material loss. Existing machine learning studies have reported encouraging results with support vector machines, k-nearest neighbours, neural networks, and ensembles. However, the predictive accuracy and reliability of these models remain constrained by limited datasets, they provide only a single corrosion rate estimate and do not account for variability inherent in operational environments. This study addresses the urgent need for more reliable, robust prediction models by consolidating a larger dataset from experimental studies and published literature. Support Vector Regression (SVR), Extra Trees (ET), and Multilayer Perceptron (MLP) neural networks regression paradigms are investigated. Using Bayesian optmization, hyperparameter tuning is conducted. This approach is essential given the strong sensitivity of nonlinear models to poorly tuned hyperparameters and the severe prediction errors that may arise from suboptimal configurations. Model performance is evaluated using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). ET achieved the highest predictive accuracy (test R2 = 0.6320; RMSE = 0.1741 mm/year; MAE = 0.0990 mm/year), effectively capturing nonlinear feature interactions while maintaining generalization under limited data conditions. In addition, Monte Carlo–based uncertainty propagation is employed to quantify prediction uncertainty arising from input variability, and data sparsity, enabling probabilistic assessment of corrosion risk. Monte Carlo simulations propagated operational and impurity uncertainties through the trained ET model, producing probabilistic corrosion distributions with P10, P50, and P90 estimates of 0.25, 0.36, and 0.47 mm/year, respectively. The proposed framework enables a shift from deterministic corrosion prediction toward probabilistic, risk-informed integrity management for SC-CO2 environment. By combining nonlinear machine learning, Bayezian hyperparameter optimization, and Monte Carlo–based uncertainty quantification, the approach provided corrosion rate predictions accompanied by confidence bounds rather than single-point estimates. This capability will support risk-based material selection, optimized corrosion allowance design, and definition of safe operating envelopes under variable impurity compositions. Furthermore, the framework can be integrated with inspection and monitoring tools to support predictive maintenance and risk-based inspection planning, reducing the likelihood of corrosion-induced failure in transport and storage systems.