Application of Machine Learning Methods in Estimating Soil Parameters from Dynamic Penetration Tests

Perhaps the static cone penetration test is one of the most popular in situ testing procedures to explore the geotechnical properties of soil layers. This device, however, cannot be easily employed in offshore sites where the soil is relatively inaccessible. On such sites, free-falling (dynamic) penetrometers are employed to provide information on the mechanical properties of the soil. These devices can provide the total time and depth of penetration as well as the deceleration characteristic of the penetrometer to infer soil properties. Nonetheless, explicit relations between the penetration characteristics and the soil properties do not exist in the literature, limiting the application of dynamic penetrometers in practice. Based on numerical methods and experimental tests, there are useful databases that relate the penetration parameters to the soil properties. In practical site investigation, the soil properties such as the shear strength, stiffness, strain rate dependency, and profile non-uniformity may be inferred from the dynamic penetration information by solving a complex inverse analysis problem. In this study, ten alternative machine learning techniques are employed to extract the soil properties from the dynamic penetration results obtained by experimental tests or numerical analysis. The results indicate that the extreme gradient boosting method outperforms its counterparts for the problem analysed in this study.