Correlations vs Machine Learning Models for Prediction of Vs From CPT: A Practice-Oriented Approach

Utilizing Cone Penetration Test (CPT) data to predict shear wave velocity (V_s) is widely practiced. This study compares V_s predictions using Machine Learning (ML) models to conventional regression for a global archive of seismic piezocone data. Prediction strategies are evaluated under grouped, nested cross-validation to emulate deployment to new soundings and sites. The database is curated into five practice-reflective feature regimes and analyzed with decision trees (DT), random forest (RF), gradient boosting (GB), support vector regression (SVR), and a shallow neural network (NN). Corrected tip resistance (q_t) emerges as the backbone predictor while pore pressure (u_2) is the highest value complement in clays. Tree ensembles attain the highest test accuracy in d+q_t+u_2 regimes. SVR and NN are competitive when u_2 is present. Steep early improvements in predictions were observed with adding clay-rich records, which plateaued beyond a few hundred samples. Relative to empirical correlations, flexible models improve R^2 by roughly 0.04 to 0.1 for feature sets that capture drainage or stress-history interactions, while regressions remain competitive in the austere q_t-only regime. We conclude through a practical lens: prioritize q_t+u_2 acquisition where feasible, expand archives when sensors are limited, and apply group-driven validation to secure transferable, bias-controlled estimates of V_s.