Abstract: The prediction of saturated-unsaturated seepage processes is crucial for the in-situ leaching of ionic rare earth ores. Currently, large-scale seepage prediction studies based on Physics-Informed Neural Networks (PINN) are relatively limited. This study constructed a long-term, large-scale neural network prediction model based on PINN for simulating in-situ leaching engineering. The model was applied to investigate the predictions of saturated-unsaturated seepage under different training sample sizes, soil types, and spatiotemporal scales. The results showed that when predicting column leaching experiments, the model achieved an R² of 0.959 when validated with experimental data and an R² of 0.98 across different data volumes and soil types in the training set. For a 10 m deep soil column, the model achieved an R² of 0.90 using only 0.4% of the simulated data for training and increased to 0.98 when the training data volume was raised to 1%, while the maximum error in the wetting front time curve was reduced to 0.36%. This model offers a novel approach and theoretical framework for predicting seepage processes in the in-situ leaching of ionic rare earth ores.