Transformer-based model for predicting jet grout column diameter
Year: 2025
Type: article
Abstract: ABSTRACT This study introduces a Transformer encoder model designed to predict the diameter of jet grout columns. The model utilizes key parameters, including Standard Penetration Test (SPT) N-values, specific energy applied during jet grouting, soil type, and jet grout type, derived from the largest compiled dataset based on published case histories and project, consisting of 239 cases. These inputs are tokenized and processed to predict jet grout column diameters. The model was trained on a training dataset, internally validated using k-fold cross-validation, and externally validated with a test dataset to enhance generalizability. The model prediction was compared against diameters calculated using empirical equations, and benchmark machine learning models including Artificial Neural Networks (ANNs), Random Forest, and XGBoost. Results demonstrate that the Transformer encoder model outperforms both empirical equations and ANNs, achieving higher prediction accuracy (R 2 =0.77) on the test dataset. While Random Forest and XGBoost showed marginally better accuracy on the test dataset, their performance on unseen design chart data was significantly weaker, highlighting the better generalization capability of Transformer encoder model. Additionally, this study employs SHAP (SHapley Additive exPlanations) for the interpretation of the Transformer encoder model. Overall, this study advances the field by introducing a Transformer encoder model for jet grout prediction, improving predictive accuracy and generalizability, and offering a structured approach for handling complex geotechnical datasets.
Source: Geodata and AI.
Authors: Rakam Lama Tamang, Yichuan Zhu, Joseph Coe
Institutions: Temple University, San Jose State University
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,000
OpenAlex ID: W4410353038