26 August 2019 in Geophysics
Full-waveform inversion (FWI) promises a high-resolution model of the earth. It is, however, a highly nonlinear inverse problem; thus, we iteratively update the subsurface model by minimizing a misfit function that measures the difference between the measured and the predicted data. The conventional l2-norm misfit function is widely used because it provides a simple, high-resolution misfit function with a sample-by-sample comparison. However, it is susceptible to local minima if the low-wavenumber components of the initial model are not accurate. Deconvolution of the predicted and measured data offers an extended space comparison, which is more global. The matching filter calculated from the deconvolution has energy focused at zero lag, like an approximated Dirac delta function, when the predicted data matches the measured one. We have introduced a framework for designing misfit functions by measuring the distance between the matching filter and a representation of the Dirac delta function using optimal transport theory. We have used the Wasserstein W2 distance, which provides us with the optimal transport between two probability distribution functions. Unlike data, the matching filter can be easily transformed to a probability distribution satisfying the requirement of the optimal transport theory. Though in one form, it admits the conventional normalized penalty applied to the nonzero-lag energy in the matching filter, the proposed misfit function is metric and extracts its form from solid mathematical foundations based on optimal transport theory. Explicitly, we can derive the adaptive waveform inversion (AWI) misfit function based on our framework, and the critical “normalization” for AWI occurs naturally per the requirement of a probability distribution. We use a modified Marmousi model and the BP salt model to verify the features of the proposed method in avoiding cycle skipping. We use the Chevron 2014 FWI benchmark data set to further highlight the effectiveness of the proposed approach.
21 October 2020 in Geophysics
Full-waveform inversion (FWI) is a nonlinear optimization problem, and a typical optimization algorithm such as the nonlinear conjugate gradient or limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) would iteratively update the model mainly along the gradient-descent direction of the misfit function or a slight modification of it. Based on the concept of meta-learning, rather than using a hand-designed optimization algorithm, we have trained the machine (represented by a neural network) to learn an optimization algorithm, entitled the “ML-descent,” and apply it in FWI. Using a recurrent neural network (RNN), we use the gradient of the misfit function as the input, and the hidden states in the RNN incorporate the history information of the gradient similar to an LBFGS algorithm. However, unlike the fixed form of the LBFGS algorithm, the machine-learning (ML) version evolves in response to the gradient. The loss function for training is formulated as a weighted summation of the L2 norm of the data residuals in the original inverse problem. As with any well-defined nonlinear inverse problem, the optimization can be locally approximated by a linear convex problem; thus, to accelerate the training, we train the neural network by minimizing randomly generated quadratic functions instead of performing time-consuming FWIs. To further improve the accuracy and robustness, we use a variational autoencoder that projects and represents the model in latent space. We use the Marmousi and the overthrust examples to demonstrate that the ML-descent method shows faster convergence and outperforms conventional optimization algorithms. The energy in the deeper part of the models can be recovered by the ML-descent even when the pseudoinverse of the Hessian is not incorporated in the FWI update.
1 October 2018 in Geophysics
We recognized over the years that our conventional surface seismic recording can effectively identify two main features of the earth: its seismic propagation attributes and those attributes resulting in echoes or reflections. Thus, the resulting expression of the earth is dominated by the generally smooth (long-wavelength) features that control wave propagation, which we use to invert for the short-wavelength features causing reflections in a process we refer to as migration velocity analysis and imaging, respectively. The features of the earth that fall in between these two model scales — the middle wavenumbers — have been elusive, which is a dilemma for full-waveform inversion because we need to build the full velocity model (a broad band of model wavenumbers). We analyze the middle model wavenumber gap, but we focus more here on potential sources of information for such middle model wavenumbers. Such sources include regularization, objective function enhancements, and multiscattered energy. Regularization, provided by a total variation (TV) constraint admits middle and high model wavenumber components into the model to enforce the model’s compliance with such a constraint. As opposed to minimizing the TV, such a constraint merely reduces the model space, and thus, these injected middle model wavenumbers are as good as the projected data information to the reduced model space. Such data information includes large offsets, but also multiscattered energy, in which the energy through wavepath and scattering updates can admit more of the elusive middle-wavenumber information that comes from the data. The combination of the right level of allowable model variations with the added data information from large offsets and multiscattering can help in filling the elusive middle model wavenumber gap and admit plausible models.
1 March 2018 in Geophysics
Macro velocity model building is important for subsequent pre-stack depth migration and full waveform inversion. Wave equation migration velocity analysis (WEMVA) utilizes the band-limited waveform to invert for the velocity. Normally, inversion would be implemented by focusing the subsurface offset common image gathers (SOCIGs). We re-examine this concept with a different perspective: In subsurface offset domain, using extended Born modeling, the recorded data can be considered as invariant with respect to the perturbation of the position of the virtual sources and velocity at the same time. A linear system connecting the perturbation of the position of those virtual sources and velocity is derived and solved subsequently by Conjugate Gradient method. In theory, the perturbation of the position of the virtual sources is given by the Rytov approximation. Thus, compared to the Born approximation, it relaxes the dependency on amplitude and makes the proposed method more applicable for real data. We demonstrate the effectiveness of the approach by applying the proposed method on both isotropic and anisotropic VTI synthetic data. A real dataset example verifies the robustness of the proposed method.
14 May 2019 in Geophysics
Cycle skipping is a severe issue in full-waveform inversion. One option to overcome it is to extend the search space to allow for data comparisons beyond the “point-to-point” subtraction. A matching filter can be computed by deconvolving the measured data from the predicted ones. If the model is correct, the resulting matching filter would be a Dirac delta function in which the energy is focused at zero lag. An optimization problem can be formulated by penalizing this matching filter departure from a Dirac delta function. Because the matching filter replaces the local sample-by-sample comparison with a global one using deconvolution, it can reduce the cycle-skipping problem. Because the matching filter is computed using the whole trace of the measured and predicted data, it is prone to unwanted crosstalk of different events. We perform the deconvolution in the Radon domain to reduce crosstalk and improve the inversion. We first transform the measured and the predicted data into the τ-p domain using the local Radon transform. We then perform deconvolution for the trace indexed by the same slope value. The main objective of the proposal is to use the slope information embedded in the Radon-transform representation to separate the events and reduce the crosstalk in the deconvolution step. As a result, the objective function tends to be more convex and stabilizes the inversion process. The result obtained for the modified Marmousi model demonstrates the proposed Radon-domain matching-filter approach can converge to a meaningful model given data without the low frequencies of less than 3 Hz and a v(z) initial model. Compared to the conventional time-space matching-filter approach, the Radon-domain approach indicates fewer artifacts in the model and better fitting of the measured data. The result corresponding to the Chevron 2014 benchmark data set also indicates the good performance of the proposed approach.
1 July 2019 in Geophysics
We have developed a method to obtain a misfit function for robust waveform inversion. In this method, called adaptive traveltime inversion (ATI), a matching filter that matches predicted data to measured data is computed. If the velocity model is relatively accurate, the resulting matching filter is close to a Dirac delta function. Its traveltime shift, which characterizes the defocusing of the matching filter, is computed by minimization of the crosscorrelation between a penalty function such as [ ] and the matching filter. ATI is constructed by minimization of the least-squares errors of the calculated traveltime shift. Further analysis indicates that the resulting traveltime shift corresponds to a first-order moment, the mean value of the resulting matching filter distribution. We extend ATI to a more general misfit function formula by computing different order moment of the resulting matching filter distribution. Choosing the penalty function in adaptive waveform inversion (AWI) as [ ], the misfit function of AWI is the second-order moment, the variance of the resulting matching filter distribution with zero mean. Because our ATI method is based on a global comparison using deconvolution, such as AWI, it can resolve the “cycle skipping” issue. We evaluate our ATI misfit function and compare it with state-of-the-art options such as least-squares inversion (L2 norm), wave-equation traveltime inversion, and AWI using schematic examples before moving to more complex examples, such as the Marmousi model. For the Marmousi model, starting with a 1D [ ] model, with data without low frequencies (no energy below 3 Hz), a meaningful estimation of the P-wave velocity model is recovered. Our ATI misfit function (first-order moment) indicates comparable performance with the AWI misfit function (the second-order moment). We also include a real data example from the Gulf of Mexico to demonstrate the effectiveness of our method.