Machine Learning Assisted Flash Calculation for Sour Gas and Crude Oil

Abstract

Abstract Compositional modeling is an effective and necessary technique for designing and optimizing complex oil and gas production processes, such as gas injection enhanced oil recovery (EOR). Generally, about 70% of computational time in compositional modeling is consumed by flash calculation. Replacing iterative flash calculations partially or completely in obtaining the satisfactorily accurate number of phases and compositionswith machine learning (ML) models is proven to be an efficient strategy to accelerate flash calculation. In this study, support vector machine (SVM), artificial neural network (ANN), and decision trees (DT) models were trained, validated, and tested based on 10,000 PVT data sets, which were artificially generated for a sour crude oil sample containing 13.32 mol% H2S and 2.9 mol% CO2. Comparative analysis showed that all three ML models could be used to make an accurate prediction. The obtained ANN model was compared with the vapor–liquid equilibrium (VLE) calculation and was used as a surrogate model for accelerating the flash calculation. The comparison revealed that the ANN model reduced the computational time of the VLE calculation by as high as 250 times. Moreover, the phase diagram of sour crude oil obtained by the ANN model was close to the phase diagram generated by VLE, which proves the accuracy and robustness of the ANN model. Overall, this work shows that ANN model is effective in accelerating flash calculation and reducing the computational time of compositional reservoir simulation.