Comparative Analysis of Traditional and Machine Learning Approaches for Estimating Undrained Shear Strength from SCPT Data

This study compares traditional empirical methods and machine learning (ML) models for estimating the undrained shear strength (su) of sensitive clays in the Saguenay region of Canada using SCPTu measurements from six sites, with penetration depths ranging from 10 to 40 m. SCPTu parameters (qt, fs, u2, and Vs) were recorded at 0.5 m intervals up to 20 m depth, while field vane shear tests (FVT) were conducted at 1 m intervals between 2 and 20 m, co-located with SCPTu points to ensure direct comparability. These paired datasets enabled the development of region-specific empirical correlations between SCPTu parameters and su, as well as the training and evaluation of multiple ML regression models. Among the six algorithms considered, Random Forest and XGBoost demonstrated the highest predictive accuracy (R2  up to 0.95), outperforming classical approaches. The findings highlight the effectiveness of integrating SCPTu data with ML techniques to enhance su estimation in overconsolidated, sensitive clays and emphasize the importance of region-focused calibration for geotechnical design.