Geotechnical Engineering and Near-Surface Geophysics—Correlations of Parameters Using Machine Learning

In certain instances, borehole data cannot easily be obtained for geotechnical engineering investigations due to environmentally sensitive soils, steep topography, or the presence of utilities. Surface geophysics methods offer an alternative approach for obtaining the desired parameters. Typical geophysical equipment for characterizing the shallow subsurface allows for measurements at sites inaccessible to a geotechnical drill rig. The resulting geophysical data can then be used as inputs to machine learning (ML) platforms to develop a mapping between the geophysical inputs and desired geotechnical parameters. Two commonly used geophysical methods are seismic refraction and electrical resistivity. We combine compressional wave refraction data, MASW data, electrical resistivity data, and induced polarization data sets as inputs to ML models that predict geotechnical parameters of choice. We have compiled a database containing geophysical and standard geotechnical parameters obtained from over 158 boreholes in seven American states over a period of several years. We used two machine learning (ML) approaches to create models mapping four geophysical measurements to a variety of geotechnical parameters. Using a multi-layer perceptron (MLP) neural network, we achieved a coefficient of determination (R2) greater than 0.7 to predict moist unit weight. Using a Support Vector Machine (SVM), we were able to predict geology types ranging from clay/silt to sand/gravel, intermediate geomaterials (IGMs), and rock, with over 70% accuracy.