Abstract: To improve the prediction of grouting volume and seepage control effects, which are difficult to assess due to the strong concealment of grouting construction, this study explores the establishment of an efficient and accurate grouting volume prediction model based on extensive on-site construction data and machine learning methods. The Gradient Boosting Trees (GBT) model is introduced to predict the unit cement consumption during the grouting process, and the Tree-Structured Parzen Estimator (TPE) algorithm is used to optimize the hyperparameters of the GBT model to enhance its prediction accuracy and generalization ability. The research results indicate that: (1) In the pre-grouting test dataset, the coefficient of determination (R2) of the TPE-GBT unit cement consumption prediction model reaches 0.80, with a mean absolute percentage error (MAPE) of 0.241 4. In the post-grouting test dataset, the model's R2 reaches 0.84, with a MAPE of 0.281 0, both of which are within an acceptable range of prediction accuracy,significantly improving the prediction accuracy compared to traditional linear regression models and GBT models;(2) Sensitivity analysis of input parameters using SHAP (Shapley Additive Explanations) values reveals that the pregrouting permeability contributes most significantly to model predictions and is a key control parameter in grouting engineering; (3) Under certain surrounding rock conditions, selecting an appropriate grouting pressure and using sequential construction methods can enhance grouting seepage control effects.