Mechanistic CO₂ Corrosion Model Optimization Supported by Machine Learning

In this work, we used a deep neural network (DNN) model to identify outliers generated by a CO₂ corrosion multi-physics (MP) model and guide the process of improving its response. Parametric studies were conducted over an extensive range of bulk pH, fluid velocity, temperature, pipe diameter and CO₂ partial pressure. The output response in terms of corrosion rate was fed into an established DNN model to identify outliers in the dataset. Considerable discrepancies were observed between the results generated by both models, particularly in the low fluid velocity regime and lower Reynolds number range. To address this, the mechanistic model was revised, updating its mesh structure and optimizing solver parameters to enhance numerical convergence and solution stability. The output from the revised mechanistic model and the DNN model were subsequently compared, the improvement in model convergence was detailed and the source of improvement was discussed. Comparative analysis of the original and revised multi-physics models revealed that the original model exhibited a bias toward underpredicting corrosion rates. This was a consequence of the relatively higher surface pH, which suppresses cathodic reactions.