ABSTRACT

Electrical-conductivity spectra of soils contain valuable information about their texture, structure, and composition that can be linked to their geotechnical properties. Concurrent measurements of electrical spectra in the frequency range of 0.01 Hz to 10 kHz and geotechnical properties, that is, the dry unit weight γd, modulus of elasticity E, and the hydraulic conductivity K, are performed on natural soil samples in a laboratory environment. The electrical spectra are modeled with the Jonscher fractal power law model characterized by three parameters: DC conductivity σDC, transition frequency fc, and an exponent n. We explore a machine-learning technique, the support vector regression (SVR) methodology, to model and predict the geotechnical properties from the Jonscher parameters, and we compare our results with the predictions of multiple linear regression (MLR). For model training and testing, the Jonscher parameters are used as the input, and a geotechnical parameter is used as the output. Model comparisons indicate that the developed SVR models predict K with an R2=0.85, predict γd with R2=0.88, and predict E with R2=0.76. In comparison, MLR models predict K with an R2=0.56, γd with R2=0.68, and E with R2=0.58. The results illustrate that the SVR models are more accurate, reliable, and achieve better performance for predicting the geotechnical properties from the electrical parameters in comparison to the predictions of the MLR models. Our study offers an opportunity in our quest in using noninvasive electrical geophysical methods to obtain geotechnical properties of soils, and it has broad implications in engineering geophysics.

You do not currently have access to this article.