Wenlong Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Wenlong Wang has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- George A. McMechan
The University of Texas at Dallas 2 , Center for Lithospheric Studies, 800 West Campbell Road (ROC21), Richardson, Texas 75080, USA.
8 shared publications - Jianwei Ma
Harbin Institute of Technology
5 shared publications - Fei Xie
Sinopec, Exploration and Production Research Institute 4 , Beijing 100083, China.
3 shared publications - Biaolong Hua
Total E&P USA Inc., Houston, USA
2 shared publications - Florian Bossmann
Department of Mathematics, Harbin Institute of Technology, Harbin, China
2 shared publications - Bertrand Duquet
Total E&P USA Inc., Houston, Texas, USA.
1 shared publication
Latest Publications
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. PP (01962892) , 1
WGAN regularized elastic full waveform inversion for petrophysical properties
In seismic exploration, Lame parameters (λ, μ) and density (ρ) are often used to discriminate litho-fluid types, while petrophysical parameters such as porosity (ϕ), clay content (C), and water saturation (Sw) provide a more detailed characterization of reservoir rock properties. The direct inversion of high-resolution rock physical profiles from seismic data is becoming popular by combining waveform inversion with rock physics modeling, especially for complex reservoirs. However, multi-parameter waveform inversions are notoriously difficult because of crosstalks between different parameters. Rock physics theories largely increase the nonlinearity of the inversion, making a direct inversion of rock physical properties extremely unstable. To alleviate the problem, we propose a regularization term utilizing positional encoded Wasserstein generative adversarial networks (WGAN-PE). The network learns multi-dimensional distributions of rock physical properties from well-log data. Positional encoding in the regularization helps the network integrate depth information with rock physical parameters and stabilize the inversion. Tests from synthetic data demonstrate that positional encoding significantly enhances the performance of WGAN regularization. Numerical experiments on three models are compared with different regularizations. The results suggest that the improved WGAN regularization gives satisfactory results for direct rock physical properties inversions.
- 3 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. PP (01962892) , 1
WGAN regularized elastic full waveform inversion for petrophysical properties
In seismic exploration, Lame parameters (λ, μ) and density (ρ) are often used to discriminate litho-fluid types, while petrophysical parameters such as porosity (ϕ), clay content (C), and water saturation (Sw) provide a more detailed characterization of reservoir rock properties. The direct inversion of high-resolution rock physical profiles from seismic data is becoming popular by combining waveform inversion with rock physics modeling, especially for complex reservoirs. However, multi-parameter waveform inversions are notoriously difficult because of crosstalks between different parameters. Rock physics theories largely increase the nonlinearity of the inversion, making a direct inversion of rock physical properties extremely unstable. To alleviate the problem, we propose a regularization term utilizing positional encoded Wasserstein generative adversarial networks (WGAN-PE). The network learns multi-dimensional distributions of rock physical properties from well-log data. Positional encoding in the regularization helps the network integrate depth information with rock physical parameters and stabilize the inversion. Tests from synthetic data demonstrate that positional encoding significantly enhances the performance of WGAN regularization. Numerical experiments on three models are compared with different regularizations. The results suggest that the improved WGAN regularization gives satisfactory results for direct rock physical properties inversions.
- 9 January 2025
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 90 (5) , V473-V484
Stability analysis of deep-learning methods for seismic data interpolation
Deep-learning algorithms are applied to solve various geophysical problems, which draw significant concerns over their stability. How to evaluate the stability of artificial neural networks has been a challenging problem for a long time. We focus on seismic data interpolation. The stability of artificial neural networks is evaluated using three criteria: tiny worst-case perturbations, small structure recovery, and hallucinations. Tiny perturbations to the input may lead to large interpolation errors. The perturbation that results in the largest error can be optimized and measured as a stability criterion. Small structure recovery measures how effectively a network can reconstruct detailed structures. Hallucinations are structures that have been falsely added by the network. The stability of three existing deep-learning algorithms for seismic data interpolation are evaluated. Quantitative analysis shows that instability is commonly observed in all networks. Different network architectures, even trained with the same data set, exhibit varied sensibility to small perturbations from the input. A UNet structure can better recover small structures within the seismic data than the other network structures. With mathematical proof and tests on numerical data, we show that an improperly prepared training set may cause hallucinations during the testing stage, especially when the network reaches a low training loss during the training stage. The developed metrics for stability analysis can provide valuable information for evaluating network performances and thus are potentially useful in other seismic data processing routines.
- 9 January 2025
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 90 (5) , V473-V484
Stability analysis of deep-learning methods for seismic data interpolation
Deep-learning algorithms are applied to solve various geophysical problems, which draw significant concerns over their stability. How to evaluate the stability of artificial neural networks has been a challenging problem for a long time. We focus on seismic data interpolation. The stability of artificial neural networks is evaluated using three criteria: tiny worst-case perturbations, small structure recovery, and hallucinations. Tiny perturbations to the input may lead to large interpolation errors. The perturbation that results in the largest error can be optimized and measured as a stability criterion. Small structure recovery measures how effectively a network can reconstruct detailed structures. Hallucinations are structures that have been falsely added by the network. The stability of three existing deep-learning algorithms for seismic data interpolation are evaluated. Quantitative analysis shows that instability is commonly observed in all networks. Different network architectures, even trained with the same data set, exhibit varied sensibility to small perturbations from the input. A UNet structure can better recover small structures within the seismic data than the other network structures. With mathematical proof and tests on numerical data, we show that an improperly prepared training set may cause hallucinations during the testing stage, especially when the network reaches a low training loss during the training stage. The developed metrics for stability analysis can provide valuable information for evaluating network performances and thus are potentially useful in other seismic data processing routines.
- 31 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-10
Toward Artifact-Free Impedance Inversion by a Semi-Supervised Network With Super-Resolution and Attention Mechanism
Estimating the subsurface impedance properties is an essential process in seismic exploration and reservoir characterization. The accuracy and efficiency of impedance inversion have been greatly improved by semi-supervised methods. However, existing semi-supervised inversion methods treat post-stack seismic traces as independent sequential time series, which causes accumulated prediction error along the time axis and horizontally noncontinuous seismic events. We propose a semi-supervised impedance inversion network. The new contribution includes two perspectives: 1. An attention mechanism is utilized to derive data-adaptive weights from both the time and positional axis, which largely reduces the artifacts in conventional semi-supervised impedance inversions; 2. A super resolution module is implemented to reconcile the dimension inconsistency between seismic data and the resultant impedance profile. By testing on the Marmousi2 model, SEAM model as well as field data, we show that the newly added modules can largely reduce the artifacts and improve the prediction accuracy for acoustic impedance.
- 31 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-10
Toward Artifact-Free Impedance Inversion by a Semi-Supervised Network With Super-Resolution and Attention Mechanism
Estimating the subsurface impedance properties is an essential process in seismic exploration and reservoir characterization. The accuracy and efficiency of impedance inversion have been greatly improved by semi-supervised methods. However, existing semi-supervised inversion methods treat post-stack seismic traces as independent sequential time series, which causes accumulated prediction error along the time axis and horizontally noncontinuous seismic events. We propose a semi-supervised impedance inversion network. The new contribution includes two perspectives: 1. An attention mechanism is utilized to derive data-adaptive weights from both the time and positional axis, which largely reduces the artifacts in conventional semi-supervised impedance inversions; 2. A super resolution module is implemented to reconcile the dimension inconsistency between seismic data and the resultant impedance profile. By testing on the Marmousi2 model, SEAM model as well as field data, we show that the newly added modules can largely reduce the artifacts and improve the prediction accuracy for acoustic impedance.
- 11 April 2023
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 88 (4) , R499-R512
Reweighted variational full-waveform inversions
Starting from an initial model and predefined priors, a variational full-waveform inversion (VFWI) seeks posterior distributions of model parameters via optimization using a Bayesian theorem. Thus, VFWI is useful in estimating uncertainties in full-waveform inversions (FWIs). However, the resolution of the inverted models from VFWIs is usually not as good as those from conventional FWIs. We decompose the loss function in a VFWI into two terms: the complexity cost and the data misfit. The data variance, which balances these two terms, is expected to decrease during an iterative optimization to gradually put higher weight on the data misfit. We develop and compare two reweighting schemes in VFWI that can produce high-resolution velocity models and the associated uncertainties. Inversions are performed efficiently in a deep-learning framework to take advantage of automatic differentiation. Tests using synthetic data and blind data indicate that reweighted VFWIs are less sensitive to initial models and can generate inversion results that have higher resolution than those from conventional FWIs.
- 11 April 2023
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 88 (4) , R499-R512
Reweighted variational full-waveform inversions
Starting from an initial model and predefined priors, a variational full-waveform inversion (VFWI) seeks posterior distributions of model parameters via optimization using a Bayesian theorem. Thus, VFWI is useful in estimating uncertainties in full-waveform inversions (FWIs). However, the resolution of the inverted models from VFWIs is usually not as good as those from conventional FWIs. We decompose the loss function in a VFWI into two terms: the complexity cost and the data misfit. The data variance, which balances these two terms, is expected to decrease during an iterative optimization to gradually put higher weight on the data misfit. We develop and compare two reweighting schemes in VFWI that can produce high-resolution velocity models and the associated uncertainties. Inversions are performed efficiently in a deep-learning framework to take advantage of automatic differentiation. Tests using synthetic data and blind data indicate that reweighted VFWIs are less sensitive to initial models and can generate inversion results that have higher resolution than those from conventional FWIs.
- 4 February 2023
- Published by Elsevier in Ultrasonics
- Vol. 131, 106930
Detection of defects in highly attenuating materials using ultrasonic least-squares reverse time migration with preconditioned stochastic gradient descent
- 4 February 2023
- Published by Elsevier in Ultrasonics
- Vol. 131, 106930
Detection of defects in highly attenuating materials using ultrasonic least-squares reverse time migration with preconditioned stochastic gradient descent
- 19 July 2021
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 86 (6) , R795-R810
Elastic isotropic and anisotropic full-waveform inversions using automatic differentiation for gradient calculations in a framework of recurrent neural networks
We have implemented multiparameter full-waveform inversions (FWIs) in the framework of recurrent neural networks in elastic isotropic and transversely isotropic media. A staggered-grid velocity-stress scheme is used to solve the first-order elastodynamic equations for forward modeling. The gradients of loss with respect to model parameters are obtained by automatic differentiation. Multiple elastic model parameters are simultaneously inverted with a minibatch optimizer. We prove the equivalency of full-batch automatic differentiation and the conventional adjoint-state method for inversions in elastic isotropic media. Synthetic tests on elastic isotropic models show that the minibatch configuration has a better convergence rate and higher inversion accuracy than full-batch elastic FWIs. Inversions with data that contain incoherent and coherent noise are tested, respectively. With automatic differentiation, we determine the ease of extension to anisotropic media with two parameterizations, and the potential to implement it for more general media.
- 19 July 2021
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 86 (6) , R795-R810
Elastic isotropic and anisotropic full-waveform inversions using automatic differentiation for gradient calculations in a framework of recurrent neural networks
We have implemented multiparameter full-waveform inversions (FWIs) in the framework of recurrent neural networks in elastic isotropic and transversely isotropic media. A staggered-grid velocity-stress scheme is used to solve the first-order elastodynamic equations for forward modeling. The gradients of loss with respect to model parameters are obtained by automatic differentiation. Multiple elastic model parameters are simultaneously inverted with a minibatch optimizer. We prove the equivalency of full-batch automatic differentiation and the conventional adjoint-state method for inversions in elastic isotropic media. Synthetic tests on elastic isotropic models show that the minibatch configuration has a better convergence rate and higher inversion accuracy than full-batch elastic FWIs. Inversions with data that contain incoherent and coherent noise are tested, respectively. With automatic differentiation, we determine the ease of extension to anisotropic media with two parameterizations, and the potential to implement it for more general media.
- 5 February 2021
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 86 (2) , U1-U13
Automatic velocity picking from semblances with a new deep-learning regression strategy: Comparison with a classification approach
The physical basis, parameterization, and assumptions involved in root-mean-square (rms) velocity estimation have not significantly changed since they were first developed. However, these three aspects are all good targets for novel application of the recent emergence of machine learning (ML). Therefore, it is useful at this time to provide a tutorial overview of two state-of-the-art ML implementations; we have designed and evaluated classification and regression neural networks for the extraction of apparent rms velocity trajectories from semblance data. Both networks share a similar end-to-end trainable structure, except for the final layer. In the classification network, the velocity picking is performed by finding the largest amplitude trajectory through all velocity bins. The regression network, on the other hand, applies a differentiable soft-argmax function that converts the feature maps directly to apparent rms velocity values as functions of traveltime. Relative confidence maps can also be estimated from both neural networks. A large number of synthetic models with horizontal layers are created, and common-midpoint gathers are simulated from those models as training samples. Transfer learning is applied to fine-tune the networks with a small number of samples for testing with synthetic and field data from more complicated (2D) models. Tests using synthetic data show that the regression and classification networks can give reasonable velocity predictions from semblances, but the regression network gives higher accuracy.
- 5 February 2021
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 86 (2) , U1-U13
Automatic velocity picking from semblances with a new deep-learning regression strategy: Comparison with a classification approach
The physical basis, parameterization, and assumptions involved in root-mean-square (rms) velocity estimation have not significantly changed since they were first developed. However, these three aspects are all good targets for novel application of the recent emergence of machine learning (ML). Therefore, it is useful at this time to provide a tutorial overview of two state-of-the-art ML implementations; we have designed and evaluated classification and regression neural networks for the extraction of apparent rms velocity trajectories from semblance data. Both networks share a similar end-to-end trainable structure, except for the final layer. In the classification network, the velocity picking is performed by finding the largest amplitude trajectory through all velocity bins. The regression network, on the other hand, applies a differentiable soft-argmax function that converts the feature maps directly to apparent rms velocity values as functions of traveltime. Relative confidence maps can also be estimated from both neural networks. A large number of synthetic models with horizontal layers are created, and common-midpoint gathers are simulated from those models as training samples. Transfer learning is applied to fine-tune the networks with a small number of samples for testing with synthetic and field data from more complicated (2D) models. Tests using synthetic data show that the regression and classification networks can give reasonable velocity predictions from semblances, but the regression network gives higher accuracy.
- 24 November 2019
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 85 (2) , U31-U46
Velocity model building in a crosswell acquisition geometry with image-trained artificial neural networks
We have developed an artificial neural network to estimate P-wave velocity models directly from prestack common-source gathers. Our network is composed of a fully connected layer set and a modified fully convolutional layer set. The parameters in the network are tuned through supervised learning to map multishot common-source gathers to velocity models. To boost the generalization ability, the network is trained on a massive data set in which the velocity models are modified from natural images that are collected from an online repository. Multishot seismic traces are simulated from those models with acoustic wave equations in a crosswell acquisition geometry. Shot gathers from different source positions are transformed as channels in the network to increase data redundancy. The training process is expensive, but it only occurs once up front. The cost for predicting velocity models is negligible once the training is complete. Different variations of the network are trained and analyzed. The trained networks indicate encouraging results for predicting velocity models from prestack seismic data that are acquired with the same geometry as in the training set.
- 24 November 2019
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 85 (2) , U31-U46
Velocity model building in a crosswell acquisition geometry with image-trained artificial neural networks
We have developed an artificial neural network to estimate P-wave velocity models directly from prestack common-source gathers. Our network is composed of a fully connected layer set and a modified fully convolutional layer set. The parameters in the network are tuned through supervised learning to map multishot common-source gathers to velocity models. To boost the generalization ability, the network is trained on a massive data set in which the velocity models are modified from natural images that are collected from an online repository. Multishot seismic traces are simulated from those models with acoustic wave equations in a crosswell acquisition geometry. Shot gathers from different source positions are transformed as channels in the network to increase data redundancy. The training process is expensive, but it only occurs once up front. The cost for predicting velocity models is negligible once the training is complete. Different variations of the network are trained and analyzed. The trained networks indicate encouraging results for predicting velocity models from prestack seismic data that are acquired with the same geometry as in the training set.
- 1 March 2019
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 84 (2) , C95-C105
Reverse time migrations in transversely isotropic media: A comparison between acoustic and elastic wave equations with two wave mode separation algorithms
Anisotropic reverse time migrations (RTMs) using pseudoacoustic or elastic wave equations are tested, and the migrated PP images, obtained using the same data set as input, are compared. In anisotropic elastic RTMs, both divergence operators, and localized low-rank approximations (LLA), are tested and compared for P/S separation. Tests with synthetic data indicate that elastic RTMs have better illumination apertures than pseudoacoustic RTMs in subsalt areas because of the involvement of converted S-waves in the former. Pseudoacoustic RTMs have diamond-shaped S-wave artifacts, which do not exist in elastic RTM images, provided that the P- and S-waves are separated in the elastic wavefields. LLAs provide affordable, accurate P/S separations in anisotropic media, and the separation results are better than those obtained using divergence operators. Anisotropic elastic RTMs with LLA give the best quality images.
- 1 March 2019
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 84 (2) , C95-C105
Reverse time migrations in transversely isotropic media: A comparison between acoustic and elastic wave equations with two wave mode separation algorithms
Anisotropic reverse time migrations (RTMs) using pseudoacoustic or elastic wave equations are tested, and the migrated PP images, obtained using the same data set as input, are compared. In anisotropic elastic RTMs, both divergence operators, and localized low-rank approximations (LLA), are tested and compared for P/S separation. Tests with synthetic data indicate that elastic RTMs have better illumination apertures than pseudoacoustic RTMs in subsalt areas because of the involvement of converted S-waves in the former. Pseudoacoustic RTMs have diamond-shaped S-wave artifacts, which do not exist in elastic RTM images, provided that the P- and S-waves are separated in the elastic wavefields. LLAs provide affordable, accurate P/S separations in anisotropic media, and the separation results are better than those obtained using divergence operators. Anisotropic elastic RTMs with LLA give the best quality images.
- 1 October 2018
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 83 (6) , S489-S503
Isotropic elastic reverse time migration using the phase- and amplitude-corrected vector P- and S-wavefields
In elastic reverse time migration (RTM), wavefield separation is an important step to remove crosstalk artifacts and improve imaging quality. State-of-the-art techniques for wavefield separation in isotropic elastic media include using the Helmholtz decomposition and introducing an auxiliary wave equation. Although these two approaches produce pure-mode vector wavefields with correct amplitudes, phases, and physical units, their computational costs are still high under current computational capability, especially for 3D large-scale problems. Based on the P- and S-wave dispersion relations, we have developed an efficient wavefield separation strategy for elastic RTM. Instead of solving a vector Poisson’s equation in the Helmholtz decomposition, we modify the phases of source wavelet as well as multicomponent records and scale the amplitudes of extrapolated wavefields with the squares of P- and S-wave velocities. This operation allows us to produce vector P- and S-wavefields with the same phases and amplitudes as the input coupled wavefields while significantly reducing computational costs. With the separated vector wavefields, we implemented a modified dot-product imaging condition for elastic RTM. In comparison with the previously proposed dot-product imaging condition, this modified imaging condition enables us to eliminate the effects of multiplication with a cosine function and hence produces migrated images with accurate amplitudes. Several 2D and 3D numerical examples are used to demonstrate the feasibility and robustness of our method for imaging complex subsurface structures.
- 1 October 2018
- Published by Society of Exploration Geophysicists in Geophysics
- Vol. 83 (6) , S489-S503
Isotropic elastic reverse time migration using the phase- and amplitude-corrected vector P- and S-wavefields
In elastic reverse time migration (RTM), wavefield separation is an important step to remove crosstalk artifacts and improve imaging quality. State-of-the-art techniques for wavefield separation in isotropic elastic media include using the Helmholtz decomposition and introducing an auxiliary wave equation. Although these two approaches produce pure-mode vector wavefields with correct amplitudes, phases, and physical units, their computational costs are still high under current computational capability, especially for 3D large-scale problems. Based on the P- and S-wave dispersion relations, we have developed an efficient wavefield separation strategy for elastic RTM. Instead of solving a vector Poisson’s equation in the Helmholtz decomposition, we modify the phases of source wavelet as well as multicomponent records and scale the amplitudes of extrapolated wavefields with the squares of P- and S-wave velocities. This operation allows us to produce vector P- and S-wavefields with the same phases and amplitudes as the input coupled wavefields while significantly reducing computational costs. With the separated vector wavefields, we implemented a modified dot-product imaging condition for elastic RTM. In comparison with the previously proposed dot-product imaging condition, this modified imaging condition enables us to eliminate the effects of multiplication with a cosine function and hence produces migrated images with accurate amplitudes. Several 2D and 3D numerical examples are used to demonstrate the feasibility and robustness of our method for imaging complex subsurface structures.