(633e) Physics- Versus Machine Learning-Based Thermodynamic Modeling: Application to Industrial Gas Treatment

Natural gas treatment and CO2 capture are important industrial processes. Solvents are typically used to remove acid gases like CO2 and H2S, mercaptans and hydrocarbons. New solvent development, as well as the optimization of existing absorption processes, require accurate modeling and simulation of the overall absorption and desorption process. One of the key thermodynamic properties for the modeling is the equilibrium between the gas and the solvent phase, i.e., Vapor-Liquid Equilibrium (VLE).

The traditional modelling approach expresses thermodynamic equilibrium between a non-ideal gas and liquid phase. Their non-ideality is modelled by a fugacity and activity coefficient, respectively modeled using Peng-Robinson and e-NRTL (electrolyte-Non-Random Two Liquid). It is based on the detailed reaction mechanism and requires prior knowledge of physico-chemical parameters of all the species in the liquid phase, and so on. This approach is quite elaborate.

A machine-learning (ML) model can be built using available experimental VLE data for the whole system, without any prior knowledge of the reaction mechanism, physico-chemical parameters, advanced thermodynamic models for non-ideal behavior, and so on. This approach might thus be much simpler.

The goal of this study was to model VLE curves using both ML and the e-NRTL-based thermodynamic model and compare their performances. We focused on the modeling of CO2-MDEA-H2O, H2S-MDEA-H2O, CO2-H2S-MDEA-H2O and CO2-Pz-H2O systems. Methyldiethanolamine (MDEA) and piperazine (Pz) are both frequently used in industrial gas treatment due to their favorable attributes, including relatively low heat of absorption and reduced corrosiveness. The developed ML models were compared to classical e-NRTL-based thermodynamic models on new experimental data: both approaches yield similar accuracies.

In a second step we modelled the VLEs for mercaptans-MDEA-H20 using machine learning. Despite the very low number of available experimental data, the results were satisfactory.

The promising approach might lead to ML-based thermodynamic models incorporated in process simulation tools.