Effect of Sampling Dataset Size and Distribution on Machine Learning-Based Surrogate Modelling of CO₂ Corrosion

Predicting CO₂ corrosion accurately using high-fidelity mechanistic models requires specialist modelling skills and is computationally demanding, both factors limiting their practical application. This paper explores the potential of using machine learning surrogates to provide corrosion predictions accurately and efficiently. Optimized models, including gradient boosted decision trees (GBDTs), deep neural networks (DNNs) and hybrid stacked ensembles, achieved high predictive accuracy (MAPE < 5%) and drastically reduced computation time. Detailed error and data requirement analyses revealed crucial insights. A strong error correlation with low Reynolds number suggests the surrogate replicates source model limitations, potentially acting as a diagnostic tool for the underlying simulation. Furthermore, data requirement analysis highlighted the benefits of optimizing the sampling strategy, with a boundary-focused dataset based on a beta sampling distribution achieving performance plateaus using significantly fewer data points than a uniform sampling strategy (~17,000 compared to ~100,000). The work demonstrates that machine learning surrogates provide significant speed-up and valuable diagnostic capabilities.