Machine learning models for fast selection of amino acids as green thermodynamic inhibitors for natural gas hydrate

Natural amino acids are non-toxic thermodynamic hydrate inhibitors without negative environmental impact, but it is difficult to accurately select the appropriate amino acid as a quick response to the operational conditions changes in the natural gas pipeline. The objective of this study was to develop mathematical models to predict the hydrate formation temperature (HFT) in presence of amino acids, capture the relationship between amino acid structure properties and their hydrate inhibition strength, and determine the optimal type and concentration to use. The HFT prediction was evaluated using multiple linear regression (MLR) and three machine learning methods including random forest (RF), M5 Rule (M5R) and support vector machine (SVM). After parameter optimization using the trial-and-error method, the coefficient of determination (R2) of the four models were 0.9328, 0.9793, 0.9795 and 0.9980, respectively. The SVM prediction of HFT outperformed other models as the root mean square error (RMSE) was 83%, 76% and 69% lower than that of the MLR, RF and M5R, respectively. Results also demonstrated that the relative importance of the amino acid concentration to the hydrate phase equilibrium was 5-fold higher than that of the intrinsic properties of the amino acid molecular. The SVM model proposed in this study served an easy-to-use tool for reliable prediction of HFT by just providing a new set of input data. This made it possible to accurately determine the minimum concentration of amino acids to be used during the gas pipeline transportation.