Polymer Controlled Oil Bank Dynamics: A Hybrid Physics-Informed Machine Learning Quantitative Framework

To address the lack of systematic quantitative characterization of oil bank dynamic evolution and unclear dominant controlling factors in polymer flooding, this study combines reservoir numerical simulation with Python-based quantitative analysis and a machine learning framework (random forest + SHAP). We established 1D and 2D reservoir models: the 1D model develops a precise quantitative characterization method for oil bank width (defined by front/rear edge saturation offsets Pf < 1.0% and Pb < 1.0%, fitted with a cubic polynomial, R2 > 0.95) and height (derived from optimal oil saturation difference time curves and integral calculation); the 2D model investigates the regulatory mechanism of reservoir heterogeneity. Based on 15,000 sets of physically consistent simulation data, the random forest model achieves high prediction accuracy (R2 = 0.98). Sensitivity analysis reveals that main flow direction permeability, reservoir temperature, and water-phase exponent (nw) of the Corey model are the dominant controlling parameters, exhibiting substantially higher sensitivity than polymer adsorption capacity and residual resistance coefficient. The oil bank height shows a negative correlation with the first two parameters, while it displays a peak-type variation with the water-phase exponent. Under heterogeneous conditions, permeability anisotropy amplifies the regulatory effect of relative permeability exponents, leading to unbalanced oil bank migration (quantified by front ratio R). This study breaks through the limitations of traditional qualitative characterization, elucidates the spatiotemporal evolution laws and heterogeneous regulatory mechanisms of the oil bank, and provides reliable theoretical and dataset support for optimizing polymer flooding schemes.