Prediction of clathrate hydrate phase equilibria using gradient boosted regression trees and deep neural networks

Abstract

Clathrate hydrate phase equilibrium is one of its most important basic physical properties, and numerous experiments and models have been studied. In this study, the gradient boosted regression tree algorithm was first applied to predict hydrate phase equilibrium conditions with multiple components in the presence of various salts, organics or pure water. In addition, a deep neural network algorithm was also used for comparison. A total of 1805 sets of experimental data were used as training sets for the modeling, and 136 sets of data were used to test the model. The input of the model includes 11 guest components, 9 kinds of salts, and 8 types of organics. Meanwhile, one representative temperature and pressure was set as the input, and the other was set as the output. The prediction results showed that the coefficients of determination (R2) of the pressure models were higher than 99.00%, the average absolute relative deviation was below 20.00%, and the average relative deviation was below 8.00%. Furthermore, the temperature model had an R2 of 99.90%, an average absolute relative deviation of 0.30% or better, and an average relative deviation of approximately 0.00%. Overall, the prediction of the temperature model was the best in this study. Furthermore, the accuracy of these models was satisfactory compared to the values in the literature.