Robust Ensemble Learning Models for Predicting Hydrogen Sulfide Solubility in Brine

Abstract

Hydrogen sulfide (H 2 S) sequestration in geological formations can be one of the promising techniques for reducing greenhouse gas emissions. Accurate predictions of phase behavior and H 2 S solubility in aqueous solution phases are vital to provide better accuracy in designing, well planning, and the process of injection well optimizations. In this study, a vast number of data sets for H 2 S solubility in pure water and aqueous solutions of NaCl have been collected. In this regard, three intelligent paradigms, including Categorical Boosting (CatBoost), Extra Trees, and Light Gradient Boosting Machine, were implemented for establishing accurate predictive paradigms of H 2 S solubility in pure water and brine. It was found that the data-driven model achieved outstanding accuracy. Among the suggested schemes, the CatBoost model outperformed the other paradigms and resulted in more accurate predictions of H 2 S solubilities at a wide range of operating pressures, temperature, and solvent salinities. In this context, the CatBoost model yielded an overall root-mean-square error of only 0.0218 and performed better than the thermodynamic-based approach. Additionally, the application of SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations methods revealed the excellent degree of explainability and interpretability of the newly proposed ensemble method for modeling the solubility of H 2 S in pure water and brine. Lastly, the newly implemented CatBoost model can help significantly in dealing with the tasks and challenges related to managing H 2 S through geological sequestration and also monitoring the issues associated with the production from sour reservoirs, mainly the monitoring of sour corrosion and controlling the rise in H 2 S content in the produced gas.