The efficient development of tight sandstone reservoirs is challenged by complex pore structures and heterogeneous remaining oil distribution. This study provides a quantitative pore-scale investigation into waterflooding mechanisms in the Changqing Oilfield’s C61 reservoir using Real Sandstone Micro-Models that preserve native pore geometry and wettability. Through integrated physical experiments and statistical analysis, we identified three displacement types (uniform, finger-like, reticulated) and found that permeability and pore-throat radius exhibit strong positive correlations with displacement efficiency (R2 > 0.75, p < 0.01), while heterogeneity indices show significant negative correlations. We further quantified diminishing returns for increased injection volume beyond two pore volumes under the tested conditions. To address the challenge of small sample size (n = 10), we employed a systematic machine learning framework incorporating data augmentation via SMOGN (Synthetic Minority Over-sampling Technique for Regression) and Gaussian noise addition, expanding the dataset to 60 samples. Four algorithms—Random Forest XGBoost, Support Vector Regression and Artificial Neural Network (Artificial Neural Network)—were evaluated using 10 × 5-fold cross-validation on the augmented dataset, with final validation on the original 10 samples using leave-one-out cross-validation (LOOCV). The Random Forest model outperformed the other algorithms under the tested conditions, achieving a cross-validation R2 of 0.85 ± 0.04, RMSE of 2.51% ± 0.42%, and MAE of 1.98% ± 0.35% in repeated 5-fold cross-validation on the augmented dataset, and a LOOCV R2 of 0.76 with RMSE of 3.18% on the original 10 samples. Feature importance analysis suggests that permeability (importance score 0.38) and pore-throat radius (0.27) are dominant controls, with the variation coefficient (0.18) indicating a detrimental role of pore structure heterogeneity. This work provides an exploratory, data-driven framework for optimizing waterflooding strategies under the tested conditions, demonstrating a potential approach to maximize insights from limited experimental data. Generalization to other reservoirs requires validation on larger, independent datasets.