14 December 2020 in Geophysics
Full-waveform inversion (FWI) is a powerful method for providing a high-resolution description of the subsurface. However, the misfit function of the conventional FWI method (metric l2-norm) is usually dominated by spurious local minima owing to its nonlinearity and ill-posedness. In addition, FWI requires intensive wavefield computation to evaluate the gradient and step length. We have considered a general inversion method using a deep neural network (DNN) for the FWI problem. This deep-learning inversion method reparameterizes physical parameters using the weights of a DNN, such that the inversion amounts to reconstructing these weights. One advantage of this deep-learning inversion method is that it can serve as an iterative regularization method, benefiting from the representation of the network. Thus, it is suitable to solve ill-posed nonlinear inverse problems. Furthermore, this method possesses good computational efficiency because it only requires first-order derivatives. In addition, it can easily be accelerated by using multiple graphics processing units and central processing units, for weight updating and forward modeling. Synthetic experiments, based on the Marmousi2, 2004 BP, and a metal ore model, are used to show the numerical performance of the deep-learning inversion method. Comprehensive comparisons with a total-variation regularized FWI are presented to show the ability of our method to recover sharp boundaries. Our numerical results indicate that this deep-learning inversion approach is effective, efficient, and can capture salient features of the model.
1 January 2017 in Computational and Mathematical Methods in Medicine
Optical tomography is an emerging and important molecular imaging modality. The aim of optical tomography is to reconstruct optical properties of human tissues. In this paper, we focus on reconstructing the absorption coefficient based on the radiative transfer equation (RTE). It is an ill-posed parameter identification problem. Regularization methods have been broadly applied to reconstruct the optical coefficients, such as the total variation (TV) regularization and the regularization. In order to better reconstruct the piecewise constant and sparse coefficient distributions, TV and norms are combined as the regularization. The forward problem is discretized with the discontinuous Galerkin method on the spatial space and the finite element method on the angular space. The minimization problem is solved by a Jacobian-based Levenberg-Marquardt type method which is equipped with a split Bregman algorithms for the regularization. We use the adjoint method to compute the Jacobian matrix which dramatically improves the computation efficiency. By comparing with the other imaging reconstruction methods based on TV and regularizations, the simulation results show the validity and efficiency of the proposed method.
23 October 2020 in Inverse Problems
The second-order derivative information plays an important role for large-scale full waveform inversion problems. However, exploiting this information requires massive computations and memory requirements. In this study, we develop two inexact Newton methods based on the Lanczos tridiagonalization process to consider the second-order derivative information. Several techniques are developed to improve the computational performance for our proposed methods. We present an effective stopping condition and implement a nonmonotone line search method. A method based on the adjoint-state method is used to efficiently compute Hessian-vector products. In addition, a diagonal preconditioner using the pseudo-Hessian matrix is employed to accelerate solving the Newton equation. Furthermore, we combine these two inexact Newton methods to improve the computational efficiency and the resolution. 2D and 3D experiments are given to demonstrate the convergence and effectiveness of our proposed methods. Numerical results indicate that, compared with the inversion methods based on the first-order derivative, both methods have good computational efficiency. Meanwhile, the method based on MINRES solver performs better than the method with Lanczos_CG due to its ability of utilizing the negative eigenvalue information when solving strongly nonlinear and ill-posed problems.
10 February 2025 in Journal of Inverse and Ill-posed Problems
Mathematically, full waveform inversion is a nonlinear and ill-posed inverse problem, requiring a regularization method to obtain a reasonable result. Total variation regularization is an effective regularization method which can preserve the sharp edges of the solution. It is well-known that the standard total variation regularization usually leads to stair-casing artifacts in slanted structures. Thus, the standard total variation regularization may not be effective in solving the full waveform inverse problem, due to the slanted properties of the subsurface structures. In this paper, we propose a rotational total variation regularization method based on the weighting rotational transform operator and the standard total variation regularization for the full waveform inverse problem. To further improve the resolution of the inverted results for different directional structures, a hybrid regularization method combining the rotational and standard total variation regularization is proposed. An efficient version of the conjugate gradient method, i.e., CGOPT, is used to efficiently solve the proposed methods. Numerical experiments based on Sigsbee and Marmousi2 models are carried out to demonstrate the effectiveness of our methods.
31 January 2026 in Journal of Computational and Applied Mathematics
28 May 2025 in Acta Geophysica