Conventional methane flooding continues to suffer significant performance limitations in heterogeneous reservoirs due to early breakthrough, mobility imbalance, and low sweep coverage, which has created the need for more intelligent and reliable approaches for predicting and optimizing gas-assisted enhanced oil recovery (EOR). This study examines methane and foam injection performance using an integrated simulation–machine-learning framework capable of forecasting reservoir response with high precision while eliminating the computational cost of repeated full-physics simulations. Numerical compositional simulation using CMG generated large production datasets (recovery factors, cumulative oil production, pressure distribution, displacement efficiency, and sweep efficiency) were used to train and validate a Random Forest Regression model (80% training, 20% testing). Results revealed contrasting recovery mechanisms: methane displacement efficiency exhibited a mean of 0.762 with a broad distribution and methane sweep efficiency averaged of 43.213%, indicating strong reservoir propagation but unstable pore-scale mobilization; whereas foam displacement efficiency averaged 0.764 with minimal spread and foam sweep efficiency averaged 27.180%, reflecting highly consistent microscopic displacement but limited macroscopic reservoir coverage. The machine-learning model achieved R² > 0.93 and RMSE values as low as 0.015, confirming high predictive reliability and demonstrating its suitability as a fast surrogate evaluation system for EOR planning. The integrated results show that methane and foam possess complementary strengths, and that the highest oil recovery in heterogeneous reservoirs was expected from hybrid injection configurations such as methane-foam co-injection or alternating slugs which balance volumetric sweep and pore-scale displacement to maximize production performance.