Classification of Cone Penetration Test data using a hybrid learning model

The classification of soils into categories with similar properties is one of the main tasks in geotechnical engineering. As this step forms the basis of the further design, it must be done in preliminary stages of a construction project, where the feasibility of the project is often not yet proved and hence, inevitable related to financial risks. As a consequence, cost and time intensive soil investigation campaigns are reduced to a minimum in order to keep the financial consequences of an unfeasible project low. Over the past years, the Cone Penetration Test (CPT) has become a cost and time effective and thus popular field test for the exploration of the subsoil conditions. Based on the almost continuous gathered data from the CPT, various soil behaviour type (SBT) charts for subsoil classifications were published (e.g., Robertson 2009, 2010, 2016). With an increasing amount of highquality data available, machine learning (ML) has found its way into various fields of research and proved to be a feasible tool for the interpretation of large datasets. In 2021, Oberhollenzer et.al. (2021) published a dataset, which consists of 1339 CPTs from Austria and southern Germany. All tests have been interpreted with the software CPTeIT of Geologismiki to determine the different soil behaviour types according to Robertson. In addition, soil classifications based on grain size distribution (EN ISO 14688) from adjacent boreholes have been assigned to 490 of these tests. These soil classes were combined into 7 classes (Oberhollenzer_classes, denoted as OC), where the soils range from gravel to clay. Based on this dataset, Rauter & Tschuchnigg (2021) performed a feasibility study to investigate the applicability of supervised machine learning for soil classification. In this study, the ability of machine learning models to classify soils from CPT data is evaluated. Furthermore, the influence of the classification (OC vs. SBT) itself on the prediction accuracy as well as the model architectures were analysed in detail. The ML models were also used to classify soils from unseen CPT data, which provided good results. However, the results presented in this paper also reveal a couple of drawbacks regarding the utilized dataset, classifications and the applied algorithms. These issues are addressed and revised in the present study.