1 January 2014 in Geophysical Journal International
Full waveform inversion (FWI) requires a hierarchical approach to maneuver the complex non-linearity associated with the problem of velocity update. In anisotropic media, the non-linearity becomes far more complex with the potential trade-off between the multiparameter description of the model. A gradient filter helps us in accessing the parts of the gradient that are suitable to combat the potential non-linearity and parameter trade-off. The filter is based on representing the gradient in the time-lag normalized domain, in which the low scattering angle of the gradient update is initially muted out in the FWI implementation, in what we may refer to as a scattering angle continuation process. The result is a low wavelength update dominated by the transmission part of the update gradient. In this case, even 10 Hz data can produce vertically near-zero wavenumber updates suitable for a background correction of the model. Relaxing the filtering at a later stage in the FWI implementation allows for smaller scattering angles to contribute higher-resolution information to the model. The benefits of the extended domain based filtering of the gradient is not only it's ability in providing low wavenumber gradients guided by the scattering angle, but also in its potential to provide gradients free of unphysical energy that may correspond to unrealistic scattering angles.
2 May 2012 in Geophysical Prospecting
30 October 2019 in Geophysics
Low-frequency seismic data are crucial for convergence of full-waveform inversion (FWI) to reliable subsurface properties. However, it is challenging to acquire field data with an appropriate signal-to-noise ratio in the low-frequency part of the spectrum. We have extrapolated low-frequency data from the respective higher frequency components of the seismic wavefield by using deep learning. Through wavenumber analysis, we find that extrapolation per shot gather has broader applicability than per-trace extrapolation. We numerically simulate marine seismic surveys for random subsurface models and train a deep convolutional neural network to derive a mapping between high and low frequencies. The trained network is then tested on sections from the BP and SEAM Phase I benchmark models. Our results indicate that we are able to recover 0.25 Hz data from the 2 to 4.5 Hz frequencies. We also determine that the extrapolated data are accurate enough for FWI application.
31 August 2021 in Geophysics
Building realistic and reliable models of the subsurface is the primary goal of seismic imaging. We have constructed an ensemble of convolutional neural networks (CNNs) to build velocity models directly from the data. Most other approaches attempt to map full data into 2D labels. We exploit the regularity of seismic acquisition and train CNNs to map gathers of neighboring common midpoints (CMPs) to vertical 1D velocity logs. This allows us to integrate well-log data into the inversion, simplify the mapping by using the 1D labels, and accommodate larger dips relative to using single CMP inputs. We dynamically generate the training data in parallel with training the CNNs, which reduces overfitting. Data generation and training of CNNs is more computationally expensive than conventional full-waveform inversion (FWI). However, once the network is trained, data sets with similar acquisition parameters can be inverted much faster than with FWI. The multiCMP CNN ensemble is tested on multiple realistic synthetic models, performs well, and was combined with FWI for even better performance.
2 September 2015 in Geophysics
The gradient of standard full-waveform inversion (FWI) attempts to map the residuals in the data to perturbations in the model. Such perturbations may include smooth background updates from the transmission components and high wavenumber updates from the reflection components. However, if we fix the reflection components using imaging, the gradient of what is referred to as reflected-waveform inversion (RWI) admits mainly transmission background-type updates. The drawback of existing RWI methods is that they lack an optimal image capable of producing reflections within the convex region of the optimization. Because the influence of velocity on the data was given mainly by its background (propagator) and perturbed (reflectivity) components, we have optimized both components simultaneously using a modified objective function. Specifically, we used an objective function that combined the data generated from a source using the background velocity, and that by the perturbed velocity through Born modeling, to fit the observed data. When the initial velocity was smooth, the data modeled from the source using the background velocity will mainly be reflection free, and most of the reflections were obtained from the image (perturbed velocity). As the background velocity becomes more accurate and can produce reflections, the role of the image will slowly diminish, and the update will be dominated by the standard FWI gradient to obtain high resolution. Because the objective function was quadratic with respect to the image, the inversion for the image was fast. To update the background velocity smoothly, we have combined different components of the gradient linearly through solving a small optimization problem. Application to the Marmousi model found that this method converged starting with a linearly increasing velocity, and with data free of frequencies below 4 Hz. Application to the 2014 Chevron Gulf of Mexico imaging challenge data set demonstrated the potential of the proposed method.
11 June 2019 in Geophysics
Obtaining high-resolution models of the earth, especially around the reservoir, is crucial to properly image and interpret the subsurface. We have developed a regularized elastic full-waveform inversion (FWI) method that uses facies as the prior information. Deep neural networks (DNNs) are trained to estimate the distribution of facies in the subsurface. Here, we use facies extracted from wells as the prior information. Seismic data, well logs, and interpreted facies have different resolution and illumination to the subsurface. Besides, a physical process, such as anelasticity in the subsurface, is often too complicated to be fully considered. Therefore, there are often no explicit formulas to connect the data coming from different geophysical surveys. A deep-learning method can find the statistically correct connection without the need to know the complex physics. In our deep-learning scheme, we specifically use it to assist the inverse problem instead of the widely used labeling task. First, we conduct an adaptive data-selection elastic FWI using the observed seismic data and obtain estimates of the subsurface, which do not need to be perfect. Then, we use the extracted facies information from the wells and force the estimated model to fit the facies by training DNNs. In this way, a list of facies is mapped to a 2D or 3D inverted model guided mainly by the structure features of the model. The multidimensional distribution of facies is used either as a regularization term or as an initial model for the next waveform inversion. Our method has two main features: (1) It applies to any kind of distribution of data samples and (2) it interpolates facies between wells guided by the structure of the estimated models. Results with synthetic and field data illustrate the benefits and limitations of this method.