Artificial neural network-based three-phase equilibria computation in compositional simulation of EOR and storage of CO2 in low-temperature reservoir

Injecting carbon dioxide (CO 2 ) into reservoirs is a widely recognized method for enhanced oil recovery (EOR) and carbon storage. This study introduces an innovative Artificial neural network (ANN)-based proxy model that significantly enhances the speed of determining equilibrium states in fluid systems, especially in the complex phase behavior of the CO 2 -hydrocarbon system. Notably, the model can handle up to three phases in isothermal compositional CO 2 injection simulations, marking a significant advancement in this field. A key innovation of this work involves developing a new, streamlined approach for generating training data tailored to capture the compositional variations typical of CO 2 injection for EOR or carbon storage. This method effectively segments the compositional range for each component, excluding the injected CO 2 , to generate representative oil samples. Pressure-composition diagrams are then generated for these samples within selected pressure intervals, and the resulting data is utilized for model training. Additionally, the ANN model incorporates a probability threshold to filter its predictions, ensuring results maintain the same standards of accuracy as traditional algorithms. This innovative model not only ensures consistency but also dramatically reduces computational time, as demonstrated through extensive numerical tests on various reservoir fluids. The ANN-based stability model showcases exceptional computational efficiency, reducing the time required to determine three-phase equilibrium status by over 95% compared to traditional methods. The model’s efficacy, evidenced by extensive testing on various reservoir fluids, highlights its potential to expedite multi-phase compositional simulations in CO 2 injection scenarios. • First machine learning-based model for vapor–liquid–liquid hydrocarbon system equilibrium. • A novel efficient and easily implemented training data generation. • Reliability threshold for model output to achieve the same results as traditional algorithm.