Machine Learning Based Surrogate Modelling of High-fidelity Multiphysics CO2 Corrosion Model Predictions

Abstract The University of Leeds has developed a high-fidelity mechanistic CO2 corrosion prediction model that simulates the underlying bulk equilibrium, mass transport, and electrochemical processes using a multiphysics simulation software. However, several characteristics of this model hinder its industry adoption, including licensing costs, the requirement for trained users, and the complexity of achieving satisfactory convergence when performing extended run-time large scale parametric sweeps. In this work we investigate the use of Machine Learning based surrogate modelling to combat these limitations, with the aim of developing a model to emulate the predictions of the existing Leeds multiphysics model which is more suitable for industry utilization. Six Machine Learning models were trained on over 160,000 CO2 corrosion rate predictions made by the Leeds multiphysics model, each with hyperparameters optimized accordingly. Preliminary analysis indicates predictions by the Deep Neural Network (DNN) and Extra Trees Regressor (ETR) models are most accurate, exhibiting the lowest root mean square error and mean absolute percentage error respectively. Further trade-offs between these models are discussed, highlighting significant improvements in simulation time, computational cost, and usability compared to the original model. Additionally, the DNN proves effective in detecting convergence issues and anomalies in Leeds multiphysics model output data.