Machine learning method for CPTu based 3D stratification of New Zealand geotechnical database sites

Three-dimensional (3D) geotechnical site stratification is of vital importance in geotechnical practice. In this study, a set of methods for 3D site stratification based on CPTu measurements of New Zealand Geotechnical Database (NZGD) sites is proposed. One-dimensional (1D) soil stratification at discrete CPTu points is first conducted and then interpolated in 3D to achieve 3D site stratification. 1D soil stratification is achieved through a proposed soil classification model combined with a proposed soil layer boundary identification method, which achieves a correct soil profile length identification rate of 93%. The soil classification machine learning model classifies the soil within NZGD into three types, i.e. Gravel, Sand, and Silt, and is able to reflect the fines content for silty sand. The model innovatively incorporates local variation information of CPTu curves in the input for a random forest algorithm to significantly improve identification accuracy to over 90%. Accurately locating soil layer boundaries is achieved through proposing a modified WTMM boundary identification method. 3D site stratification is then realized through 3D interpolation of 1D stratification at discrete CPTu points using a generalized regression neural network (GRNN) method. The 3D site stratification method is validated for two independent geotechnical sites within NZGD, exhibiting the effectiveness of the proposed set of methods.