Application of extreme gradient boosting for predicting standard penetration test N-values from cone penetration test data

Abstract

The standard penetration test (SPT) and cone penetration test (CPT) are widely used for subsurface stratigraphic characterization and the determination of geotechnical properties. While the CPT is increasingly adopted in site investigations, the SPT N-value continues to be a key parameter for developing empirical design formulas. As a result, establishing an accurate correlation between CPT data and SPT N-values remains a critical challenge in geotechnical engineering. In this study, a new side-by-side SPT-CPT database is constructed to evaluate existing conventional models, revealing their shortcomings and highlighting the need for a more reliable model. An innovative approach using an Extreme Gradient Boosting (XGBoost) model to predict SPT N-values from CPT data. This model overcomes the limitations of conventional transformation models by leveraging machine learning algorithm that can capture complex relationships within the data. The new XGBoost model incorporates a broader range of input variables compared to conventional models, including cone resistance, sleeve friction, soil behavior type index, fines content, depth of CPT data, and effective overburden stress. Through comparative analyses with other prevalent machine learning models, including random forests, back-propagation artificial neural networks, and support vector machines, we demonstrate that the XGBoost model significantly outperforms both conventional and machine learning-based models in terms of accuracy and robustness.