Deep learning for estimating porous material properties from full-waveform data

Abstract

This research focuses on both developing computational tools to model transient wave propagation in coupled poroviscoelastic-viscoelastic-acoustic media and estimating material properties from the recorded full-waveform data. Numerical simulation of wave-dominated problems is computationally demanding. Efficient parallelization, capability to handle complex geometries, and sufficient numerical accuracy are some of the requirements for a suitable full-waveform simulation technique. On the other hand, the robustness and prediction accuracy are needed from the method used to solve the corresponding inverse problem. In this work, the discontinuous Galerkin method is used to solve the forward model while the convolutional neural networks is used to solve the estimation problem. Two-dimensional model problems with simulated data are presented. In the numerical experiments, the primary unknowns are estimated while the remaining parameters which are of less interest are successfully marginalized.This research focuses on both developing computational tools to model transient wave propagation in coupled poroviscoelastic-viscoelastic-acoustic media and estimating material properties from the recorded full-waveform data. Numerical simulation of wave-dominated problems is computationally demanding. Efficient parallelization, capability to handle complex geometries, and sufficient numerical accuracy are some of the requirements for a suitable full-waveform simulation technique. On the other hand, the robustness and prediction accuracy are needed from the method used to solve the corresponding inverse problem. In this work, the discontinuous Galerkin method is used to solve the forward model while the convolutional neural networks is used to solve the estimation problem. Two-dimensional model problems with simulated data are presented. In the numerical experiments, the primary unknowns are estimated while the remaining parameters which are of less interest are successfully marginalized.