DATA-DRIVEN MODELING OF UNDRAINED SHEAR STRENGTH IN SOFT DELTAIC CLAYS: EVIDENCE FROM THE NIGER DELTA, NIGERIA

ABSTRACT

This study investigates the application of machine learning (ML) algorithms for predicting the undrained shear strength (Su) of soft deltaic clays in the Niger Delta, Nigeria. Four supervised learning models, Linear Regression (LR), Support Vector Regression (SVR), Decision Tree Regression (DTR), and Random Forest Regression (RFR)—were trained using soil index properties, including natural moisture content, Atterberg limits, and unit weights, as predictors. Model performance was evaluated with mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²) for both training and testing datasets. Results show that predictive performance is strongly site-dependent. At Escravos and Abigborodo Sapele, SVR with polynomial kernels achieved the best balance between accuracy and generalisation, highlighting its ability to capture nonlinear soil–strength interactions. In contrast, tree-based ensemble methods, particularly RFR, provided superior results at Eremor, confirming their robustness in more heterogeneous contexts. At Cawthorne Channel, however, none of the models generalised effectively, illustrating the limitations of purely data-driven approaches under conditions of extreme heterogeneity and restricted input diversity. This challenge echoes broader findings in geotechnical ML research, where overfitting and instability are common in noisy or data-sparse environments. These findings suggest that incorporating hybrid or physics-informed ML approaches could reduce spurious correlations and embed soil mechanics principles into predictive frameworks. Three contributions emerge: model choice must remain context-specific, as no single algorithm performs consistently across sites; ensemble methods are robust but still vulnerable to overfitting; and embedding geotechnical knowledge into ML frameworks offers a pathway toward more reliable prediction in deltaic soils, where depositional variability challenges conventional design.

 

Keywords: undrained shear strength; machine learning; Niger Delta; soft clays; regression models; Random Forest