Utilizing undisturbed soil sampling approach to predict elastic modulus of cohesive soils: a Gaussian process regression model

This study addresses a critical issue of sample disturbance in predicting the elastic modulus (Es) of soft cohesive soils using machine learning techniques. Traditional approaches either use inaccurately disturbed soil samples or require complex in-situ testing, making accurate predictions of Es challenging. We aim to develop high-performing prediction models for Es of soft cohesive soils using undisturbed soil samples and gaussian process regression (GPR). To achieve this, a new laboratory dataset is established using undisturbed cohesive soil samples obtained through a Shelby tube sampler. GPR-based prediction models are then developed, with Es as the output parameter and six index properties as input features. These features include sand content (S), fine content (FC), liquid limit (LL), plastic limit (PL), water content (w), and soil density (d). The input parameters are organized into four groups (Group-1: S, FC, LL, PL, w, and d; Group-2: S, FC, LL, PL, w; Group-3: S, FC, LL, PL; Group-4: FC, LL, PL) for the development of four GPR-based models. Results indicate that a model incorporating all six input features demonstrates excellent performance, with R2 (correlation coefficient), RMSE (root mean square error), and MAE (mean absolute error) values of 0.999, 0.054, and 0.042, respectively. The effectiveness of the optimal model is further validated using field cone penetration test (CPT) data. Moreover, sensitivity and parametric investigations reveal that soil density, water content, and Atterberg limits significantly influence the characterization of Es in cohesive soils.