Abstract: Accurate groutability prediction is essential not only for mine water-inrush prevention, but also for reducing excessive cement consumption and the associated carbon footprint of grouting operations. However, field geological datasets are often small, which limits the reliability and generalization of data-driven models. A geology-informed prediction framework is presented for predicting unit grouting amount (uga) from three routinely measured borehole variables: water inflow rate, hydraulic pressure, and groundwater level. A generative augmentation strategy was employed to expand the training data from 44 to 880 samples, and the input variables were organized into a physically informed feature sequence. Based on this representation, a hybrid deep-learning model integrating a bidirectional temporal convolutional network, a bidirectional gated recurrent unit, and an attention mechanism was developed, with its hyperparameters optimized using the Crested Porcupine Optimizer. The results demonstrate that data augmentation improved the test performance, with the coefficient of determination (R²) increasing from 0.8851 to 0.9293 and the root mean square error (RMSE) reduced by 21.6%, effectively alleviating overfitting. External validation yielded R² values of 0.9982, 0.9884, and 0.9272 under identical, similar, and distinct geological settings, respectively. Distribution-shift analysis further highlighted the importance of hydrogeological-mechanism consistency for external generalization. Interpretability analysis using SHapley Additive exPlanations and Accumulated Local Effects indicated that borehole water inflow rate was the dominant predictor, contributing 75.4% to the model output, while high unit grouting amounts were associated with jointly elevated fracture connectivity and hydraulic pressure. The proposed framework provides practical support for accurate uga prediction and material-efficient grouting in intelligent mining.