Artificial neural networks in predicting of the gas molecular diffusion coefficient

Abstract

The diffusion coefficient is one of the most important parameters for designing two-phase operations between liquid and gas phases in refineries and petrochemical industries, as well as for the gas injection process in oil fields to enhance oil production. Accurate knowledge of this parameter is essential for the prediction of the dissolution rate of the gas phase into the liquid phase. Ideally, this parameter should be obtained experimentally. Given that setting up laboratory equipment and conducting experiments can be costly and time-consuming, mathematical modeling is used as an alternative. In some cases, this data is not either available or reliable, which poses a challenge for designs. Hence, empirically derived correlations are used to predict molecular diffusion coefficients. However, the success of empirical models depends mainly on the range of data at which they were originally developed. Empirical models are not comprehensive for applying to the other data. Recent studies demonstrated that the alternative approach to modeling complex processes and identifying the effective parameters is the artificial neural network (ANN), a suitable prediction method. This study presents a new model developed to predict the molecular diffusion coefficient of methane in crude oil. The model is developed using 172 data points collected from recent literature. Out of the total laboratory data, 90% (155 data points) were used for training the desired neural network, while 10% (17 data points) were reserved for testing and evaluating the performance of the network. The multi-layer perceptron (MLP) neural network architecture with back-propagation (BP) training algorithm was used successfully for the prediction of diffusion coefficients of methane in crude oil. The developed model is compared with the empirical data, which shows the developed model predicts the methane molecular diffusion coefficient in crude oil with an average absolute error of 4.18%.