Practical application of machine learning on fast phase equilibrium calculations in compositional reservoir simulations

Abstract To accurately describe the fluid phase behaviour in reservoir simulation, Equation-of-State-based compositional models are usually used. However, phase equilibrium calculations, including stability tests and phase splitting calculations, may require huge computational costs. An improved artificial neural network model is developed based on our previous work to achieve the prediction of phase stability with high accuracy and reduce the computational costs in order of magnitude. This model does not directly tell if a hydrocarbon mixture at given compositions, pressure and temperature is stable or unstable, and it is able to predict the saturation pressures. By comparing the given pressure with the predicted saturation pressure, it is natural to tell the stability of a hydrocarbon mixture. For the phase splitting calculations, another artificial neural network model is developed to provide more reliable initial guesses for commonly used equilibrium methods to reduce their nonlinear iterations. Compared with our previous work, multi-layer neural networks are employed in the models. The improved models have more efficient training processes and more accurate prediction results. Additionally, an adaptive data generation process is introduced to optimize the quality and size of training data; the data transformation and normalization are discussed to reduce the skewness of the data and rescale the data; and a model training and selection strategy is also discussed to efficiently train a high-quality model. The efficiency of the ANN models is first validated by standalone phase equilibrium calculations with a three-component hydrocarbon mixture. Then the ANN models are successfully applied in compositional simulation examples, which demonstrates the practical value of these ANN models on speeding up the phase equilibrium calculations during compositional simulations.