Identification of Soil Strata from In-Situ Test Data Using Machine Learning

An increasing number of geotechnical analyses such as settlement and stability calculations are performed using 2D or 3D models. In order to obtain a representation of the soil conditions that is as close to reality as possible, subsurface investigations such as core drillings or soundings are necessary. Based on the findings of these investigation campaigns, similar soils are grouped into layers and subsequently ground models are defined, which serve as the basis for conventional analysis or numerical investigations. In recent years, the Cone Penetration Test (CPT) has established itself as a sound alternative to costly and time-consuming core drillings. However, when using CPT, the subsoil is not visible to the engineer, making the interpretation of CPT data difficult. For the classification of soils into groups with similar properties, various soil behavior type charts based on the CPT data were published. Additionally, data science tools such as machine learning are increasingly being used to interpret CPT data. However, regardless of the interpretation method used, defining the layer boundaries and combining them into a three-dimensional model is usually done manually, thus it is still based on experience. The present work aims to provide an automatic identification of soil layers from CPT data using a combination of machine learning models. The presented approach was also tested using CPT data from a benchmark site in Austria and showed very good agreement with results of traditional (experience based) interpretation of soil investigations.