Prediction model for grouting volume using borehole image features and explainable artificial intelligence

Abstract

Grouting is widely used in various engineering projects, and accurate prediction of grout volume is of significant importance. Machine learning methods have been extensively applied in grout volume prediction, yet existing research still has limitations, such as the underutilization of unstructured data like images and the reliance on black-box models that lack interpretability and transparency, hindering practical application. This study collected field trial data from Yunnan Province, China, and introduced borehole image data, which had been overlooked in previous studies, along with explainable machine learning (XAI) methods to construct grout volume prediction models with high accuracy and interpretability. The results show that integrating image features into the dataset significantly enhanced the model's interpretability, prediction accuracy, and stability. The ensemble model, particularly Gradient Boosting Machine (GBM), achieved a Coefficient of Determination (R²) of 0.848, demonstrating strong explanatory power. SHAP analysis facilitated model interpretation and knowledge discovery, especially regarding the interactions between geological structural features and other features. The main contributions of this study are: (1) introduced and verified the validity of image data, expanding the research resources; (2) deepened the research in this field based on knowledge mining by interpretable machine learning; (3) provided new technical support for the actual grouting operation, as well as methodological references for subsequent research.