<p>Conventional deterministic models for earth-rock dam seepage prediction typically cannot satisfy both accuracy and interpretability requirements, owing to neglected hysteresis effects and empirical parameter dependence. To overcome these limitations, a mechanistic model was developed by integrating: (1) normal distribution-based hysteresis functions, (2) Polar Light Optimization (PLO) algorithm, and (3) radial basis function neural network (RBFNN) surrogate modeling. The framework involved: first, developing a surrogate model through PLO-optimized RBFNN hyperparameters using finite element results; second, a normal distribution-based hysteresis effect function is introduced to characterize dynamic seepage responses, with PLO automatically optimizing hysteresis parameters to establish a multivariate regression equation. Validation using measured data from six monitoring points in the main and auxiliary dams demonstrates that the model accurately quantifies hysteresis parameters, with results consistent with the permeability characteristics of dam materials and the physical principles governing monitoring point locations, which are unobtainable by conventional deterministic models. For points exhibiting significant hysteresis, the prediction accuracy is slightly improved (within 2%) compared to conventional deterministic models, while for non-hysteretic points, the model maintains accuracy while reducing the number of modeling factors, thereby mitigating overfitting risks. Residual analysis shows that the ± 2σ deviation coverage reaches 94.8–97.1%. The proposed approach successfully combines physical interpretability with engineering practicality, providing a novel methodology for safety assessment of earth-rock dam seepage.</p>