A New Phase-Labeling Method Based on Machine Learning for CO2 Applications

Abstract

AbstractPhase labeling can be very challenging for complicated compositional simulation cases. Inaccurate labeling can lead to issues ranging from incorrect resource accounting to non-convergent simulation runs. Accurate phase labeling algorithms are computationally demanding and are seldom used in commercial workflows. Instead, cheaper but inaccurate empirical methods are employed such as the Li-correlation (Reid et. el. 1966).Phase labelling based on critical temperature alone mis-identifies fluids below the dew point pressure as liquids rather than vapour. This is a particular problem when performing surface flashes of CO2 or H2S rich fluids since both components have critical temperatures above standard temperature. This can lead to failures in the well model, for example when a well is controlled by gas rate but the produced phase is identified as a liquid. The second part of this paper therefore describes a new phase labeling method that uses both the critical temperature and saturation pressure predictions from the ML models to generate accurate labels. Results are presented for CO2 rich fluids. We show that this ML approach can result in accurate labeling and can outperform traditional methods in computational efficiency. We also show the application on simulation cases with complicated field management scenarios that require accurate phase labeling at the in-situ as well as separator conditions.The ML workflow is based on a set of two interacting fully connected neural networks, one a classifier and the other a regressor, that are used to replace physical algorithms for single phase labelling and improve the convergence of the simulator. We generate real time compositional training data using different mixing strategies between the injected and the in-situ fluid compositions that can exhibit temporal evolution. In many complicated scenarios, a physical critical temperature as well as saturation pressure does not exist, and the iterative sequence fails to converge. We train the classifier to identify, a-priori, if a sequence of iterations will diverge. The regressor is then trained to predict an accurate value of critical temperature and saturation pressure. A framework is developed inside the simulator based on TensorFlow that aids real time machine learning applications. The training data is generated within the simulator at the beginning of the simulation run and the ML models are trained on this data while the simulator is running. All the run-times presented in this paper include the time taken to generate the training data and train the models.