(651b) A Hybrid CO2/O2/H2s Corrosion Modeling Approach Using Machine Learning and CFD Velocity Data for General Pipe Configurations

Machine learning (ML) of CFD predicted CO₂ corrosion model have been demonstrated in our earlier publications using CFD as the baseline tool to predict the maximum CO₂ corrosion rates in pipes with bends [1][2][3]. However, it was found that further inclusion of other corrosion mechanisms requires a very large amount of CFD simulations deemed unfeasible in costs and time due to the coupling of CO₂/O₂/H₂S corrosion kinetics [4], which makes the required number of parameters increased from 7 to 9. The increase does not seem significant but the required amount of the CFD simulations grows exponentially.

However, a careful examination of the CFD data in a full 3-D analysis shows that most of the surface chemistry is only dependent of the bulk chemistry data, i.e., the average species concentrations. This is due to the fact the surface corrosion chemistry occurs only near the wall surface and cannot be resolved by the CFD mesh. In fact, in CFD analysis the corrosion rate was approximated by a combination of near-wall mass transfer limitations based on the location variation of velocity, and chemical species concentrations. It was found that while the velocity local variation remains significant, the local variations of the resolved bulk (i.e. free stream) chemical species concentrations are negligible, therefore, the true surface corrosion rates are only impacted by the mass transfer limitations in most scenarios.

Nonetheless, there are scenarios where the local concentration variations cannot be ignored. In a pipe configuration such scenarios only occur when the surface corrosion rate is in the same order of the local mass transfer rate. This only occurs when the velocity is vey slow and the mass transfer rate near the wall is insufficient to remove the chemical species concentration gradient created by the surface reaction thus allows the gradient to be smoothed out by molecular diffusion. However, such process can only exist in laminar flows and is of little interest for the corrosion applications discussed here where the highest Damkohler number is only around 0.1. With this observation, the ML model is only required to predict the local velocity data and then combined with the corrosion kinetics developed by Nesic et al. [4][5] This approach greatly simplifies the CFD data generation process and allows a hybrid CO₂/O₂/H₂S ML model to be developed.

In addition to the benefits of accelerated CFD data generation, this approach also enables further optimization of the Sherwood number-based mass transfer model. By using reinforcement learning algorithms the constant and exponents in the Sherwood number (CReⁿSc^m) can be further optimized for better accuracy. The findings will be summarized and reported.

References

[1] Kuochen Tsai, “Corrosion modeling using electrochemistry and computational fluid dynamics”, 2017 AIChE Annual Meeting, Minneapolis, MN, Oct. 29 – Nov. 3, Paper No. 335b.

[2] Yang, Huihui , Lu, Ligang , and Kuochen Tsai. "Machine Learning Based Predictive Models for CO₂ Corrosion in Pipelines with Various Bending Angles." Paper presented at the SPE Annual Technical Conference and Exhibition, Virtual, October 2020. doi: https://doi.org/10.2118/201275-MS.

[3] Kuochen Tsai, Huihui Yang and Ligang Lu, “Acceleration of CFD based corrosion models using Machine Learning Algorithms”, Paper 67a, Nov. 16-20, 2020 Virtual AIChE Annual Meeting, San Francisco, USA.

[4] S. Nesic, J. Postlethwaite, and S. Olsen. “An Electrochemical Model for Prediction of Corrosion of Mild Steel in Aqueous Carbon Dioxide Solutions”. CORROSION (1996) 52 (4): 280–294. https://doi.org/10.5006/1.3293640.

[5] S. Nesic, H. Li, J. Huang and D. Sormaz, “An open source mechanistic model for CO2/H2S Corrosion of carbon steel”, Paper No. 09572, NACE Corrosion Conference & Expo, 2009.