Application of Soft Computing Techniques for Prediction of Seismic Wave Velocity

Determination of primary wave velocity
$(V_{p})$
and shear wave velocity
$(V_{s})$
requires highly skilled professionals, costly and sensitive instruments, and specific site conditions for conducting such tests. Hence, industries and researchers often rely on predictive models to minimise the geotechnical investigation procedure. However, most of the popular regression models are site-specific. Therefore, a database of 482 samples has been created to develop AI models through experimental investigation at local sites and from published literature. The input parameters are Standard Penetration Test
$(N)$
value, depth, bulk density and site class. The seismic refraction test has been conducted for obtaining the primary wave velocity at a local site to develop a Multi-variate regression (MVR) model. Feedforward Artificial neural network (ANN) and Support vector regression (SVR) models were developed in the present study. The SVR models have been developed using Gaussian Radial Basis Function (GRBF), Exponential radial basis function (ERBF), polynomial and linear kernels. The performances of the AI models were compared with that of traditional MVR models. The study demonstrates that developing a global intelligent predictive model for predicting
$V_{p}$
and
$V_{s}$
at any site is possible with accuracy much higher than conventional MVR models. It is observed that SVR outperforms ANN and MVR models for the prediction of seismic wave velocities.