Thermodynamic features-driven machine learning-based predictions of clathrate hydrate equilibria in the presence of electrolytes

Abstract

Gas hydrates have significant applications in the areas of natural gas storage, desalination and gas separation. Knowledge of the thermodynamic conditions associated with hydrate formation is critical to their synthesis. Presently, we use machine learning (ML) to train and evaluate the performance of three algorithms on an experimental database (>1800 data points) to predict hydrate dissociation temperatures as a function of the constituent hydrate precursors and inhibitors. Importantly, and in contrast to most previous studies, we use thermodynamic variables such as the activity-based contribution due to electrolytes, partial pressure of individual gases, and specific gravity of the overall mixture as input features in the prediction algorithms. Using such features results in more physics-aware ML algorithms, which can capture the individual contributions of gases and electrolytes in a more fundamental manner. Three ML algorithms, Random Forest (RF), Extra Trees (ET), and Extreme Gradient Boosting (XGBoost) are employed and demonstrate excellent accuracy in their predictions of hydrate equilibrium conditions. The overall coefficient of determination (R2) percentage is greater than 97% for all the ML models. XGBoost outperforms RF and ET with the highest overall coefficient of determination (R2) and the lowest overall Average Absolute relative deviation (AARD) of 99.56% and 0.086% respectively.