Insights into predicting equilibrium conditions of clathrate hydrates of methane + water-soluble hydrate former

Abstract

This study aims to improve the prediction of equilibrium conditions in methane hydrate systems by incorporating diverse water-soluble hydrate formers and applying advanced machine learning techniques. Methane hydrates, which naturally form under high pressure and low temperature, can be more efficiently formed or dissociated by altering thermodynamic conditions using these hydrate formers. Accurate prediction of these conditions is crucial for optimizing gas storage and energy applications. In this research, molecular descriptors and operational parameters, such as mole fraction and pressure, are used as input variables to predict equilibrium temperature. Machine learning methods, including Decision Trees (DT), Random Forests (RF), Support Vector Machines (SVM), and Multi-Layer Perceptron (MLP), were employed with a novel data-splitting approach based on hydrate formers rather than traditional sample-based methods. Among these models, the RF achieved the highest performance, with a coefficient of determination (R²) of 0.930, a root mean square error (RMSE) of 1.71, and an average absolute relative deviation (AARD) of 0.48%. Feature selection, preprocessing, and Shapley Additive Explanations (SHAP) provided valuable insights into the influence of specific variables on model predictions. Additionally, a supplementary examination, termed the reduced model, highlights the critical role of proper feature selection, with certain features regarded as less important yet essential for the functionality of distance-based models, particularly for models like SVM and MLP. This work advances methane hydrate research by offering a more accurate and interpretable framework for predicting hydrate equilibrium, addressing key gaps in previous studies, and extending its applicability to a broader range of systems. Moreover, the introduction of a former-based data-splitting method improves generalization across different hydrate formers, while the use of SHAP values for model interpretability offers deeper insights into the relationships between molecular descriptors and hydrate equilibrium conditions. This study paves the way for improved selection of hydrate formers in hydrate systems.