(67a) Acceleration of CFD Based Corrosion Models Using Machine Learning Algorithms

A computational fluid dynamics (CFD) model combined with electrochemistry has been developed to prediction CO₂ corrosion in pipelines [1]. The CFD model can explicitly account for corrosion chemistry with parameters of temperature, CO₂ concentration, pipe geometry and local fluid velocity. This allows a full parametric simulation of the maximum corrosion rate on steel pipes with arbitrary geometries. For general applications to work, commonly used pipe shapes with bends can greatly benefit the industry that allows real-time results for CO₂ corrosion predictions. However, the idea can be also extended to other geometrical shapes for various equipment and processing units prone to corrosion. Further extension to O₂ and H₂S corrosion, and multiphase flows can be highly beneficial as well. In this work, we introduce a real-time machine-learning (ML) model that replicates the CFD predicted maximum corrosion rate in pipes with various bend angles, pipe inner diameter and bend radius within a range of flow conditions. A total of 7 variables, including pipe ID, flow inlet velocity, pH value, CO₂ partial pressure, pipe bend angle, pipe bend radius and temperature, are parametrized to generate the maximum corrosion rate using CFD. A total of 83,160 simulation cases were performed using 7 ANSYS Fluent licenses with 48 CPU cores each, and 7 weeks of computational time.

A light gradient boosting (LGBM) ML algorithm [2] has been applied to model the CFD data. The optimized model gave a resulting R²value of 0.992 and a value of 0.975 for another new set of CFD data, indicating high efficiency of the model. The forward inferencing of the LGBM model only takes a few milliseconds on a laptop computer in comparison to the 10 minutes computing time of the CFD simulation using 48 CPU cores. The completed model was implemented in an interactive web tool where users can obtain results with graphics instantaneously. One interesting observation is that the web results can be more accurate than CFD simulations because the errors resulted from incorrect problem setup or numerical divergence can be avoided since the ML model always provides a locally filtered outcome that smoothes out potential erroneous results.

The value of this work lies in the wide range of applicability for general piping corrosion with CO₂. The proposed method can be extended to other corrosion mechanisms as well, e.g., O₂ and H₂S. However, corrosion is a complex phenomenon and results can vary significantly due to metal types, surface deposits and corrosion product films. Nevertheless, this method can be integrated with other more complex corrosion mechanisms and aid the scale-up or down of experiments or plant observations that cannot be easily obtained otherwise. Some examples will be provided and discussed, and other possible extensions will be addressed.

References:

[1] Kuochen Tsai, “Corrosion modeling using electrochemistry and computational fluid dynamics”, 2017 AIChE Annual Meeting, Oct 31, Minneapolis, MN, USA, Paper No. 355b.

[2] Cheng Chen, Qingmei Zhang. QinMacBinYu, “LightGBM-PPI: Predicting protein-protein interactions through LightGBM with multi-information fusion”, Chemometrics Intell. Lab. Syst., vol. 191, pp. 54-64,2019.