Background and Objective Minimum miscibility pressure (MMP) is a key parameter for injection-pressure design in miscible gas flooding. Conventional empirical correlations and tabular machine learning models provide limited descriptions of the coupled relationships among reservoir temperature, crude-oil composition, injection-gas composition, and heavy-end properties, often leading to large prediction deviations under complex gas-source conditions.

Methods Based on 710 gas-flooding MMP measurements collected from published literature, this study developed an MMP prediction model coupling Adaptive Frequency Enhancement Transformer (AFEformer)-based feature-domain enhancement with histogram-based gradient boosting regression (HGBR). Twenty-five basic variables were arranged in a fixed order to form a within-sample feature sequence, providing a consistent mapping of cross-sectional variables for feature extraction without implying temporal dependence. Following sample-wise normalization, frequency-domain enhancement, and trend-residual decomposition, the original, enhanced, and residual features were fused into a 75-dimensional model input. The evaluated models included extreme gradient boosting (XGBoost), extremely randomized trees (Extra Trees), support vector regression with a radial basis function kernel (SVR-RBF), a deep MLP with hyperbolic-tangent activation (Deep MLP Tanh), ridge regression, and HGBR using only the original features.

Results Results for 105 independent test samples showed that the AFEformer-HGBR model achieved a coefficient of determination (R²) of 0.9640, a root mean square error (RMSE) of 1.6071 MPa, and a maximum absolute error (MaxAE) of 4.8067 MPa, outperforming all evaluated models. Compared with the original-feature HGBR, R² increased by 0.0302, RMSE decreased by 26.2%, and MaxAE dropped from 14.2581 to 4.8067 MPa. The substantial reduction in RMSE and MaxAE indicates that the fused features improve overall prediction accuracy while strengthening the control of large deviations. In particular, the decrease in MaxAE shows that the model compresses the upper end of the error distribution rather than merely improving average predictive performance.

Significance and Implications This characteristic is important for preliminary injection-pressure selection because a large error in an individual sample may lead to an unsuitable estimate of the required miscibility pressure. Permutation importance and Shapley additive explanations (SHAP) identified reservoir temperature and injection-gas composition as the main information sources for MMP prediction. Their contributions are consistent with the effects of temperature and injected-gas components on phase behavior, component exchange, and multicomponent mass transfer during miscibility development. The model retains the physical meaning of the original variables while supplementing information on variable combinations and local deviations. By reducing the influence of large prediction deviations on preliminary injection-pressure selection, the model can narrow the pressure range requiring experimental verification and improve the efficiency of MMP evaluation under complex gas-source conditions.