Solving of the Inversion Problem of the Method Multichannel Analysis of Surface Waves Based on the Artificial Neural Network

Abstract

The study is devoted to improving the efficiency of the method of multichannel analysis of surface waves due to the implementation of a new algorithm for the selection of 1D-models of the shear wave velocity by inverse dispersion curves based on machine learning. It is proposed to use a multilayer fully-connected artificial neural network (ANN) for the estimate of horizontally layered models composed of interpolated velocity values on a uniform grid with an arbitrary step. Various types of correlation dependence of the parameters of the velocity model on dispersion curves were investigated. The optimal type of the objective function was determined for ANN learning on the basis of these studies. The advantages of using a trained ANN for the problem of inversion dispersion curves are: the ability to estimate layered models with sufficient accuracy for their use at subsequent stages of seismic data processing, resistance to noise, no significant requirements for computational resources and the absence of the need for additional adjustment of parameters. The tests of the trained ANN, with the architecture proposed in this study, show high accuracy in validation (about 96%), which indicates the success of training and the actually perspective of using neural networks for inversion problems.

An artificial neural network approach for the inversion of surface wave dispersion curves

“…From this point of view, the fine‐layer approach seems more reasonable. In our previous research, we tried to apply the ANN method to invert dispersion data for a fine‐layered model composed of 50 layers with a fixed depth step (Yablokov et al ., 2019). We computed synthetic dispersion data for a set of models composed of two to five layers and tried to process it by ANN.…”