Interpretable machine learning-based modelling of minimum miscibility pressure in hydrocarbon gas injection processes

Abstract Gas injection is a highly effective method for enhancing oil recovery from hydrocarbon reservoirs by improving displacement efficiency and reducing residual oil saturation. The minimum miscibility pressure (MMP) is an essential characteristic that governs the gas injection process in enhanced oil recovery. This research aimed to estimate the MMP of gaseous hydrocarbon-crude oil systems employing artificial neural networks. To this end, a comprehensive dataset comprising 135 experimental data points, covering a wide range of temperatures and gas compositions, was used to train and validate four intelligent models, namely radial basis function (RBF), multilayer perceptron (MLP), generalized regression neural network (GRNN), and cascade forward neural network (CFNN), optimized with advanced optimization algorithms. The model inputs included the reservoir temperature, the mean critical temperature of injected gas, the molecular weight of the C 5+ component ( $$\:{\text{M}\text{W}}_{{\text{C}}_{5+}}$$ ), and the mole percentage of intermediate (H 2 S, CO 2 , and C 2 -C 4 ) and volatile (N 2 , C 1 ) components. The findings revealed that the GRNN achieved the highest predictive performance for MMP, yielding an average absolute percent relative error (AAPRE) of 4.74%. Trend analysis illustrated that an increase in reservoir temperature leads to a higher MMP values, while higher mean critical temperature of the injected gas results in a reduction of MMP. Furthermore, sensitivity evaluation was conducted to pinpoint the dominant parameters influencing MMP. The results underscored that the $$\:{\text{M}\text{W}}_{{\text{C}}_{5+}}$$ has the most pronounced positive effect, indicating that heavier hydrocarbon components elevate MMP, while the mean critical temperature of the injected gas emerged as the most significant factor in reducing MMP. To enhance the interpretability of the developed machine learning models, SHAP (SHapley Additive exPlanations) analysis was conducted. The results indicated that Tc, ave–gas (K) exhibited the most extensive spread of SHAP values, highlighting its dominant role in driving output variability. Also, according to the outlier detection analysis, more than 99% of the data fall within the valid range. The developed approach provides a reliable and computationally efficient tool for screening and optimizing gas injection processes in oil reservoirs.