Machine Learning accelerated Phase Flash Calculation for CO2–Brine System considering Capillarity Effect

Summary The geological sequestration of carbon dioxide (CO2) has been proposed as a critical solution to mitigate climate change. Accurate prediction of CO2 solubility in brine is necessary for a better estimate of CO2 storage capacity in subsurface reservoirs. Specific Equation of State (EOS) models coupled in compositional reservoir simulators are usually used to predict the fate of dissolved CO2 over time. However, this approach can be computationally expensive, particularly if a more detailed physics description is used, such as capillarity, which is important in describing phase behavior in the cap-rock layers. To address this issue, a Machine Learning (ML) based approach is used to generate a proxy from high-fidelity physics simulations describing CO2-brine phase behavior. First, Peng-Robinson EOS, coupled with Duan & Sun model are used to generate approximately 5000 samples of CO2 solubility calculation dataset for ML model training. Several ML models such as Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN) are then trained and compared to predict phase split calculation result and CO2 solubility. Potential ML model is then optimized by hyper-parameter tuning to increase prediction accuracy. Sensitivity analysis from the trained model was performed to evaluate the behavior of model prediction. The trained DNN model shows promising potential to accelerate Pressure-Volume-Temperature (PVT) and solubility calculations, with a test case showing approximately a 210 times speedup with an Average Absolute Percentage Error (AAPE) of less than 0.117% compared to the EOS calculation.