Uncertainty quantification of phase velocity surface waves multy-modal inversion using machine learning

Abstract

The paper is devoted to uncertainty quantification of the inverse problem solution of the multichannel analysis of surface waves method - the inversion of the curves of the phase velocity via frequency dependence. The uncertainty estimation approach is based on the Monte Carlo sampling strategy and a multilayer fully connected artificial neural network to approximate nonlinear dependence of shear wave velocity and layers thickness via values of phase velocity surface waves. Frequency-dependent noise in the data and errors of the inverse operator are projected onto the inverse problem solution. The results of unimodal and multimodal inversion are compared on the example of synthetic data processing. The experimental results show that using of machine learning approaches makes it possible to quickly and accurately estimate the posterior probability density of the reconstructed velocity model parameters.