Water-soluble organic former selection for methane hydrates by supervised machine learning

An explainable supervised machine learning was used to inspect the insight into the former selection for methane hydrate forming systems in the presence of water-soluble organic molecules. The former is an important ingredient that allows methane hydrate formation at conditions closer to the ambient in solidified natural gas technology (SNG). Over 800 samples were collected from literatures and experiments and utilized for supervised modeling. The data were split into train and test sets with an 80:20 ratio, preprocessed, and used to build predictive models, which are linear regression-based and tree-based models. Categorical Boosting Machine Regressor (CatBoost) performed the best on the equilibrium temperature prediction using 10 relevant attributes on the unseen data set with R2 of 0.973, RMSE of 1.375, and MAPE of 0.269 %. Feature selection suggested that only 7 necessary attributes are adequate to make a model, which is comparable to the full model. The model was then explained via Shapley Additive Explanations (SHAP) analyses. Cyclic molecules without hydrogen bond donors, with low polarity, at a mole fraction in the range of 0.02 - 0.07 were suggested to be effective formers that can shift the equilibrium temperature to higher levels.(c) 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Automated summary

In this study, explainable supervised machine learning was used to examine the influence of water-soluble organic molecules on methane hydrate forming systems. With over 800 samples, predictive models were built using CatBoost, which yielded good temperature predictions on unseen data. Feature selection showed that only 7 attributes are necessary for an effective model.