A high-fidelity mechanistic carbon steel/carbon dioxide (CO2) corrosion prediction model developed at the University of Leeds has been implemented in the multi-physics (MP) software package COMSOL. The model integrates the bulk solution equilibria, interfacial mass transport and solution reactions, and surface anodic and cathodic electrochemical processes.

In this work, the high-fidelity model is randomly sampled in the 5-input dimensions (bulk pH, temperature, CO2 partial pressure, piping velocity and diameter) using a Beta distribution, B(α, β), with the parameters α and β set to 12 in order to oversample the extremes of the input range. The high-fidelity MP model was sampled approximately 225,000 times and the data refined using a Deep Neural Network (DNN) as described by Proudlove et al.. The output from the high-fidelity model went through three rounds of refinement via the DNN to produce the final dataset.

After data refinement and 5-fold cross-validation, the data was further analyzed using a symbolic regression (SR) genetic algorithm. Each generation from the SR was plotted on a Pareto optimization plot, displaying the relationship between model complexity and the fit quality metric, (1 − R2). From the Pareto front, showing the trade-off between maximum quality of fit for minimal model complexity, the ‘best’ mathematical expression representing the data was chosen. Additionally, the mathematical expressions generated were reviewed subjectively for ‘physical reasonableness’ and for ‘ease of implementation in a typical spreadsheet. The output from the three models were compared in a 3-dimensional correlation diagram with excellent agreement between all models, R2 > 0.99.

Subject
Correlation, Modeling, Corrosion rate, Machinery, Structures, Tools, Probability, Algorithms, Statistics, Mathematical models, Corrosion modeling, Computer based testing, Risk reduction
Keywords:
Symbolic Regression, Deep Neural Network, Machine Learning, Carbon Dioxide Corrosion, Modeling, Simulation
© 2025 Association for Materials Protection and Performance (AMPP). All rights reserved. This work is protected by both domestic and international copyright laws. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means (electronic, mechanical, photocopying, recording, or otherwise) without the prior written permission of AMPP. Positions and opinions advanced in this work are those of the author(s) and not necessarily those of AMPP. Responsibility for the content of the work lies solely with the author(s).
2025
Association for Materials Protection and Performance (AMPP)
You do not currently have access to this content.