Optimizing Minimum Miscibility Pressure Prediction Using Machine Learning: A Comprehensive Evaluation and Validation

This study provides the proof-of-concept for identifying the most suitable machine-learning (ML) model that predicts minimum miscibility pressure (MMP) based on temperature, crude oil, and injected fluid composition. MMP defined as the lowest pressure injected gas developing miscibility with reservoir oil is crucial for gas-enhanced oil recovery. Slimtube experiments considered the most reliable for MMP predictions are time-consuming. Although researchers have considered ML to expedite MMP predictions, validation of the optimal model that integrates the main controlling factors remains outstanding. We tested eight ML models of different complexities to determine the most suitable for predicting MMP. The models were trained and tested using 75 and 25% of 142 publicly available slim-tube experiments and validated using six in-house slim-tube MMP experiments. The injected gas compositions varied and included H₂S, CO₂, N₂, CH₄, and C₂⁺. We assessed model suitability using mean absolute error (MAE). Models with MAEs <7% estimated the MMP. The highest-performing model after testing and validation was the neural network. This work identifies the most suitable machine-learning technique for MMP prediction validated using recent experiments. The optimal model provides an instantaneous MMP chart for gas injections typical to a Permian field. Also, we demonstrate a workflow for recommending optimal injection gas compositions with low MMP and reduced emissions associated with gas EOR. The procedure ultimately reduces the cost of performing the slim-tube experiment.