Surface Wave Inversion by Neural Networks (Radial Basis Functions) for Engineering Applications

The shear properties of soils are important in the engineering and design of foundations and soil embankment structures. Conventional inversion of surface waves for these shear properties is typically done by inverting measured dispersion curves. Unfortunately, the measurement of a phase velocity dispersion curve is subject to a number of uncertainties. These include spatial aliasing and the misidentification of modes which may occur from a limited recording window (time and space). We present an alternative scheme which maps seismograms directly to velocity profiles (pattern recognition). Our method only requires two geophones and is immune to spatial aliasing. No identification of modes is necessary. All that is required is that the field data parameters match those of the synthetics used to train the network. The method is intended as a means to extend borehole control by surface measurements. One application would be real time inversions of surface waves in the field on a personal lap-top computer. Included is a field data example which successfully demonstrates the technique. Finally, the new neural net procedure described here will be of value in many other geophysical settings where interpolation of irregularly spaced data is needed.