Predicting undrained shear strength of remolded fine-grained soils using traditional and machine learning models

Abstract Critical parameters needed for geotechnical design are frequently estimated using empirical correlations derived from laboratory classification and strength tests. These tests are essential techniques for assessing the engineering behaviour of cohesive soils. Since undrained shear strength and consistency limits are important engineering characteristics that describe soil behaviour, it is crucial to evaluate the reliability of these parameters obtained from different testing techniques. In this study, the physical and index properties of 40 cohesive soils with varying characteristics were determined in the laboratory. For each soil, five samples with different water contents were prepared and subjected to Casagrande, fall cone, and laboratory vane shear tests. The undrained shear strength values obtained from vane shear tests were used as reference to evaluate the variation of the fall cone factor for each soil. Based on the liquid limit values obtained from the fall cone and Casagrande tests, undrained shear strength was estimated using liquidity index and water content ratio parameters. The undrained shear strength predicted from fall cone data exhibited a high level of accuracy, achieving R 2 ≈ 0.89. Additionally, similarities and differences between the models were analysed by comparing the single-variable equations created in this study with empirical correlations found in the literature. ANN, RF, SVM, XGB, and stacking machine learning models were used in addition to traditional statistical methods to forecast the undrained shear strength. Overall, the machine learning framework demonstrated superior predictive performance, and the top-performing model, RF, demonstrated reliable estimation capability for undrained shear strength, with R 2 ≈ 0.97.