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

The University of Leeds has developed a high-fidelity mechanistic CO₂ corrosion model that simulates the underlying bulk equilibrium, mass transport, and electrochemical processes using multiphysics simulation software. This high-fidelity model (the Leeds Model) provides accurate forecasting of corrosion rate (CR), surface pH, and the precipitation of protective films within mild steel pipelines. However, there are a number of barriers hindering the effective use of the outputs from this model, including the requirement for trained users, licensing costs and the timescale and complexity of achieving satisfactory convergence when performing extended run-time, large scale parametric sweeps.

In this work we investigate the use of Machine Learning (ML) based surrogate modelling to combat these limitations, with the aim of developing a surrogate model to emulate Leeds Model predictions in a much more convenient format. Six ML models were trained on over 160,000 CO₂ CR predictions made by the Leeds Model, each with hyperparameters optimised accordingly.

Of the models examined, the Deep Neural Network (DNN) and Extra Trees Regressor (ETR) exhibit most accurate test set predictions, achieving a MAPE of 0.44% and 0.39% respectively. Further trade-offs between these two surrogates are discussed, highlighting significant improvements in simulation time, computational cost, and usability compared to the original model. During analysis, 47 anomalous datapoints are discovered within the training dataset, caused by poor convergence in the Leeds Model. DNN predictions are shown to be insensitive to poorly converged datapoints and prove effective in detecting convergence issues in the Leeds Model’s CR predictions.