25 November 2024 in Geophysical Journal International
SUMMARY: Full waveform inversion (FWI) of ground-penetrating radar (GPR) data is one of the key methods for reconstructing high-resolution near-surface images and understanding changes in the near-surface environment. However, FWI faces great challenges due to the ill-posed nature of the optimization problem and its high sensitivity to the quality of the initial model. In this paper, we propose the implicit full waveform inversion (IFWI) algorithm for GPR that addresses these limitations by using an implicit and continuous function built by deep neural networks to represent multiple subsurface parameters. In this way, IFWI can be thought of as training a physically constrained neural network representation, allowing it to exploit the frequency principle in deep learning optimization, that is first recovering the low-frequency components of the subsurface parameters and large-scale structures, and then their high-frequency and fine-grained information, which makes IFWI an automated multiscale inversion approach without manual frequency band selection. Numerical experiments are conducted to verify the effectiveness and performance of IFWI relative to FWI. The results demonstrate that IFWI can accurately reconstruct subsurface images and anomalies even with a suboptimal initial model, while FWI and multiscale FWI struggle to provide useful information due to being trapped in local minima. Consequently, IFWI exhibits less dependence on the initial model compared to FWI and multiscale FWI. To better understand the behaviour of IFWI and its sensitivity to neural network architecture, we comprehensively discuss the influences of these hyperparameters on IFWI, providing initial selection guidelines and practical guidance for adjustments.
1 February 2023 in Geophysics
Full-waveform inversion (FWI) aims at estimating subsurface physical parameters by minimizing the misfit between simulated data and observations. FWI relies heavily on an accurate initial model and is less robust to measurement noise and physical assumptions in modeling. Compared with FWI, wavefield reconstruction inversion (WRI) is more robust to these uncertainties but faces high computational costs. To overcome these challenges, we have developed a new form of WRI. This reformulation takes the form of a traditional FWI formula, which includes a medium-dependent weight function, and can be easily incorporated into the current FWI workflow. This weight function contains the covariance matrices to characterize the distribution of uncertainties in measurements and physical assumptions. We discuss various options of the theoretical covariance matrix of the new inversion method and find how they relate to various well-known approaches, including FWI, WRI, and extended FWI. On the basis of the preceding comparison, we develop a theoretical covariance matrix definition based on the source. Numerical experiments demonstrate that our method with a source-dependent theoretical covariance matrix is more computationally efficient than conventional WRI while preserving a certain degree of robustness.
24 November 2022 in IEEE Transactions on Geoscience and Remote Sensing
Full Waveform Inversion (FWI) methods evaluate subsurface media properties by minimizing the misfit between synthetic and observed data. However, its derivation process omits measurement errors and physical assumptions in modeling, resulting in many problems in practical utilizations. Wavefield Reconstruction Inversion (WRI) can handle non-physical data by combining theoretical and measurement errors but is computationally expensive in large-scale or 3D cases. An alternative formula for WRI in the form of traditional FWI (WRI in the FWI form, FWRI) was developed. The new form includes a medium-dependent weight function or a data domain Hessian matrix. Based on the data domain Hessian matrix, a feasible accurate time-domain solution is provided for 2D and 3D FWRI. Specifically, the point spread function (PSF) method and an-isotropic TV regularization were used to calculate the weighted residuals. Furthermore, an approximate solution based on the forward operator is provided, considering the efficient application in 3D. Numerical tests using the homogeneous model show that accurately computed residuals mitigate the effects of "cycle-skipping" and non-physical data. Sensitivity kernel analysis demonstrates the low wavenumber update mechanism of FWRI and verifies the robustness of the data domain Hessian to "cycle-skipping". The robustness of the proposed algorithm is demonstrated using a 2D Marmousi model with partially elastic perturbations and a visco-acoustic Red Sea model. Furthermore, the 3D Overthrust model is solved with an approximate solution to demonstrate the feasibility of the proposed method in 3D applications.
31 October 2022 in Geophysics
Target-oriented inversion (TOI) can resolve subsurface reservoir parameters based on surface waveform data. Moreover, virtual data at the datum level and a precise local forward operator are required to accurately perform inversion in the target area. The Marchenko redatuming, a state-of-the-art seismic redatuming method, is used to obtain the virtual data of the target area. Local forward operators of crosscorrelation and cross-convolution forms are used to generate the local wavefields. A subsurface-domain interferometric objective function and the corresponding inversion algorithm are developed. A simple model is used to demonstrate the accuracy of this approach. This technique is applied to the Chevron 2014 benchmark data set to invert a specific area, indicating that the method can achieve accurate high-resolution inversion results with reduced computational costs. Finally, a time-lapse inversion is performed to exhibit the practicality of the new approach in 4D seismic exploration. These numerical results demonstrate great applications of the TOI based on the redatumed data and local forward operators of crosscorrelation or convolution forms.
28 April 2025 in IEEE Transactions on Geoscience and Remote Sensing
Seismic full waveform inversion (FWI) is a powerful technique for generating high-resolution images of the subsurface, leveraging the rich information content within recorded seismic waveforms. However, due to certain limitations in data acquisition and processing, the optimization of seismic FWI is inherently non-linear and non-unique, implying the possibility of multiple solutions that can adequately account for the observed data. Unlike deterministic inversion, which seeks a single best-fit solution, probabilistic inversion explores a range of subsurface model parameters adhering to probability distributions that fit the observations within a given confidence level. In this paper, we introduce a variational autoencoder (VAE)-based probabilistic FWI method to assess the uncertainty associated with subsurface parameters using unsupervised deep learning. By repeatedly sampling the latent representation distributions of the observed data, a set of predicted results that adhering to the subsurface model’s posterior distribution can be reconstructed using the decoder and then calculated for uncertainty quantification. Compared to conventional Markov Chain Monte Carlo-based approaches that require substantial computational costs beyond deterministic FWI, the proposed VAE-based FWI offers the opportunity to assess the uncertainty without additional computational demands. Furthermore, the VAE-based method is benchmarked against other deep learning-based approaches, including AE-based deterministic FWI and dropout-based probabilistic FWI, using 2-D Marmousi and Overthrust models. The comparison reveals that the VAE-based approach offers a more nuanced and comprehensive inversion result compared to the single prediction from the AE-based method. Furthermore, it effectively mitigates the dropout-based method’s tendency to underestimate uncertainty in deeper layers, offering a more reliable assessment of inversion uncertainty.
30 September 2022 in IEEE Transactions on Geoscience and Remote Sensing
Wavefield reconstruction inversion (WRI) is a novel seismic inversion method with a more expansive search space and reduced nonlinearity and is more robust than other conventional techniques [e.g., full-waveform inversion (FWI)] in terms of local minima. Most of the modifications and applications of WRI are generally used in the frequency domain, while the current time-domain solutions for WRI are either based on rough approximations, which lead to inaccurate inversion results, or computationally expensive due to the data-domain Hessian. Here, we will provide another perspective to address time-domain WRI and maintain its effectiveness, namely, reducing nonlinearity and the effects of “cycle-skipping” while ignoring the data-domain Hessian. On the basis of iteratively solving the inversion problem, the time-domain augmented wave equation is solved by approximating the reconstructed wavefield on the right as the current wavefield. Due to this approximation and the traditional $L^{2}$ measurement, the reconstructed wavefield based on the poor model will be inaccurate away from the receivers, which will affect the subsequent model inversion. To alleviate this problem, we incorporate the unbalanced optimal transmission (UOT) distance measurement into the inversion. This distance measurement method can accurately obtain the best misfit between two unbalanced signals. Moreover, the accurately measured misfit with low frequency and long wavelength acts as an extended source of the subsequent wavefield reconstruction, ensuring the reconstructed wavefield’s accuracy based on the defective initial model without excessive additional computation. The numerical results demonstrate the accuracy and applicability of the proposed method.