Forge ahead in acquisition and model building via full-waveform inversion: Gulf of Mexico case study Open Access

Accurate earth models play a significant role in seismic imaging. Despite many years of migration algorithm advancement, the image could still be out of focus and seismic events lose their continuity if the model were inaccurate. In the past few decades, 3D full-waveform inversion (FWI) with an acoustic approximation of the two-way wave propagation has been widely used to build detailed earth models. The technology was pioneered by Lailly (1983), Tarantola (1984), and Pratt and Shipp (1999). However, due to the lack of sufficient computer power, the 3D version of FWI has not taken off until the time of the 21st century. The early successful examples of FWI applied to 3D data sets include the implementation in the frequency (Sirgue et al., 2008) and the time domain (Vigh and William Starr, 2008). After early success, 3D FWI has been performed on real data sets in marine, mostly towed-streamer and ocean-bottom node (OBN) acquisitions (Plessix, 2009; Sirgue et al., 2010; Vigh et al., 2010), and in land (Plessix et al., 2010; Murphy et al., 2021) environments. These examples showed that FWI can be used for velocity updates if the acquired data provide long offsets in the range of 8–25 km and low frequencies in the range of 1.5–4 Hz. In these early experiments, mostly the diving waves and early arrivals were used in conjunction with a relatively good starting model to avoid issues of cycle skipping.

With the early success in FWI especially in shallow gas-related geology, the industry’s interest persists in the potential for automated subsurface model building, especially in high-velocity contrast environments such as imaging targets associated with salt, carbonate, or igneous rocks, through the data-driven residual minimization used by FWI (Guasch et al., 2019). Significant business impact to hydrocarbon exploration is inferred through improved turnaround time in identifying targets by reducing the overall number of manual model reconstructions that are routinely performed in seismic data processing to generate sufficient quality images. By itself, FWI has proven versatile as a model building tool in the past few decades by improving the accuracy and accelerating turnaround time for earth models.

In recent years, information from the full seismic record in the seismic experiment has started to be included in FWI. All types of seismic waves are involved in the optimization scheme, including diving waves, reflections, and multiscattering waves such as prismatic waves, multiples, and, in some examples, S waves. Although the wavefield can be captured with different seismic acquisition methods such as narrow-azimuth (NAZ), wide-azimuth (WAZ), and full-azimuth (FAZ) multivessel circular shooting, recent OBN acquisition geometries prevail due to their FAZ and long-offset characteristics and the ability of OBN acquisition to record frequency content as low as 1.6 Hz. However, because OBN data collection is expensive, the technology was not used for regional studies until the node spacing was relaxed (Dellinger et al., 2017), allowing economical coverage of more than 1000 km² areas with ultralong offsets, thanks to an extensive source patch. The sparse node design has a few important components, such as offset length, node spacing, and the lowest frequency generated by the source. Recent designs are focusing on how FWI can build a velocity model only without using the data for imaging too (Blanch et al., 2020; Xing et al., 2020). To increase the productivity of the sparse node surveys, simultaneous source shooting that has a rich history in the industry is used (Cheng and Sacchi, 2014; Singh et al., 2016; Li et al., 2019). The questions are how good the starting model must be for the deep part of the velocity field to explain the data without cycle skipping and how reliable the FWI update is in the model space based on a given acquisition geometry. These questions can be investigated by perturbing the existing model and checking the FWI diving wave illumination with the planned or already executed acquisition geometry (Ahmed, 2018; Fagua Duarte et al., 2023).

Even though legacy streamer acquired data are extensively used in FWI velocity model building, the newly acquired sparse OBN data give significant uplift to the velocity model. Recently, FWI technology has evolved to use the entire seismic record, not only the early arrivals in the FWI to update the models. The advancement in the FWI algorithm, along with workflow tailored to the acquisition, can obtain the most out of the data, which improves the seismic image through FWI model updates. In spite of the improvement, the industry turned to the long offset and low-frequency acquisition to further enhance the images with the FWI-updated models.

FWI starting models are mostly mature legacy models in the Gulf of Mexico (GOM); however, the salt geometries that are extracted mostly from legacy streamer data may still have significant data misfit on the ultralong-offset OBN data. The traditional model building practice for salt-related geology is to use tomography to build the sediment-only velocity model. After developing the background sediment model, a sediment flood and, subsequently, a salt flood are performed to interpret the top of salt and then the base of salt for most of the salt structure. The salt geometry construction becomes more complex when there are salt overhangs that must be integrated into the model. To build the complex salt shape, numerous salt scenario trials may be needed to possibly delineate the salt and there is a chance that the data-fitting FWI encounters cycle skipping in spite of the effort, even though the recorded frequencies are lower and lower.

Recent FWI improvements loosened the constraints on the starting model by emphasizing the long-wavelength component of the FWI gradient (Jiao et al., 2015; Engquist et al., 2016; Warner and Guasch, 2016). Once the large-scale kinematic error was reduced, the FWI workflow was then switched to the least-squares type of objective function to further refine the model. The latest efforts involve mixing kinematic and dynamic updates in one objective function with a weighting scheme to decide the relative contribution of these two terms. One example of such improvement, the so-called enhanced-template-matching (ETM) objective function was further improved by Cheng et al. (2023) to address complex geology for long-offset data. The new objective function allows FWI to accurately update high-contrast boundaries in the models, in the complex geology, foremost among them, salt and carbonate and the deep Mesozoic delineation. The sparse node OBN unleashed the power of FWI and revealed geologic features that have never been seen before (Williams et al., 2019; Roende et al., 2020; Lin et al., 2021; Vigh et al., 2021b) and the next logical step is to pursue the elastic wave equation to simulate the data for the forward modeling and obtain more of the earth model parameters related to the shear wavefield. The obvious reason would be that the observed data are elastic and when acoustic approximation is assumed in FWI, amplitude discrepancy could occur at high-velocity contrast interfaces, such as salt boundaries, resulting in an incorrect velocity update around and inside the salt. We demonstrate that when equipped with an elastic propagator, the elastic FWI (EFWI) can achieve better model updates than its acoustic version.

A least-squares (L2-norm) objective function is used in the conventional FWI to calculate the difference between the observed data and simulated data. The known cycle-skipping issue makes this least-squares misfit function difficult to use in complex geology when low frequencies are not present in the acquired data with a sufficient signal-to-noise ratio (S/N), i.e., the objective function is exposed to the half-wavelength convergence criteria which are determined based on the lowest frequency in the observed data. Evaluation of the starting model is a challenging task because the long-wavelength component and the salt geometry need to be interrogated for possible cycle skipping which can occur in localized areas. This condition makes FWI laborious to execute especially for large project sizes. In practice, the starting model is often unable to predict data that match the acquired data to the accuracy of better than one-half wavelength of the lowest frequency available in the field data. In this circumstance, the conventional FWI with the least-squares objective function will be trapped in the local minima. Introducing an objective function that can mitigate cycle skipping is therefore an important part of FWI.

Several authors proposed explicitly deriving phase- or traveltime-difference-based objective functions. Bozdağ et al. (2011) introduce an objective function based on instantaneous phase difference, which makes the signal constant inherently lower than the acquired data. Shah et al. (2012) and Alkhalifah and Choi (2012) propose unwrapped phase-based inversions that focus on phase differences. Métivier et al. (2018) use the optimal transport distance to minimize the misfit between the observation and prediction. Alternatively, Warner and Guasch (2016) mitigate cycle skipping by deriving matching filters between the synthetic and recorded data and applying them to one of the two data sets. A few other methods to address cycle skipping and amplitude discrepancy include an extensive list such as Luo and Schuster (1991) using the traveltime information and Ma and Hale (2013) suggesting warping the synthetic data to the observed data to take care of the cycle skipping. In the industry, the traveltime-based objective function is widely used, as the traveltime difference of seismic events is more linear related to the velocity error than the waveform deviations, and any disagreement in the amplitude difference is not as important. The time shift or matching filter derivations may not reflect the correct mismatch because they are one dimension, so lateral shifts, which are particularly important in complex geology, are not considered. These erroneous time shifts could lead to incorrect velocity updates when salt or high-velocity contrast is present in the geology. Therefore, further development was implemented to improve the traveltime-based objective function in FWI, referred to as ETM-FWI (Vigh et al., 2019), which matches the local templates between the acquired and simulated shot record for temporal and spatial shift depending upon the data sampling.

The temporal shift can be resolved in a 2D or 3D mode depending on the acquired data sampling to ensure proper matching of events between the acquired and simulated data in complex geologic settings. FWI can directly minimize temporal and spatial shifts computed as local attributes in the ETM objective function by backprojecting the local shifts into the model space. After mitigating the cycle-skipping problem, the amplitude term will play a major role in the objective function to match the waveform itself. When we compare the 1D traveltime objective functions to the one that considers spatial shifts as well, the model update shows better geologic consistency enabled by better data fitting. ETM-FWI is designed to reveal the accurate salt geometry with a given input model as long as sufficiently low frequencies and sufficiently long offsets have been recorded. The ETM objective function has been proven to be less sensitive to the cycle-skipping problem than the traditional least-squares objective function while retaining the ability to reshape complex salt geometry.

In addition to the selection of the objective function, due to the geometric spreading and attenuation in the data and the modeling, it is important to take amplitude decay into account to compensate for the model update from shallow to deep depth. Ideally, a Hessian matrix could be computed for conditioning the gradient. However, the Hessian matrix calculation is usually prohibitively expensive. Practically, we use an approximation rather than directly calculating the full Hessian matrix. Such approximation only accounts for the diagonal of the Hessian (Virieux and Operto, 2009).

The other aspect of FWI is to maximize the effectiveness of the individual updates for each inversion iteration to keep the required number of iterations small because FWI is computationally costly. Traditional optimization methods for FWI in exploration geophysics are gradient-based methods such as steepest descent and nonlinear conjugate gradient (NLCG), which are computationally attractive for large-scale inverse problems. Often, the inverse of the Hessian can be replaced by a scalar, the so-called step length. The step length can be estimated using line search for which a linear approximation of the forward problem is used. For the steepest descent method, the updates may not take the fastest route to the global minima, but it can be very robust even if the data are very noisy, and it could be a viable solution against the more often used NLCG with line-search method. More advanced optimization methods, such as limited-memory Broyden-Fletcher-Goldfarb-Shanno or Newton-type methods (Pratt et al., 1998), are also frequently used but they usually require very good quality gradients to succeed and arrive at reliable velocity updates.

In many sedimentary basins around the world, prolific oil reservoirs are commonly covered by a very complex overburden, such as large canopies of salt or very complex salt bodies in the GOM, heterogeneous salt bodies in offshore Brazil, or complex layers of evaporites and clastic sediments with very high velocities in the Red Sea and the Gulf of Suez. The main challenge for imaging these reservoirs under complex overburdens was related to nonideal reservoir illumination, due to the limited offsets and azimuths of the earlier NAZ towed-streamer acquisition geometries, which were used routinely in 3D seismic acquisition since 1982 when 3D marine seismic acquisition started. The limited offsets and single azimuth in NAZ acquisition do not fully illuminate the reservoirs and thus they do not enable the derivation of an accurate velocity model for imaging using the acquired seismic data. Consequently, the reservoirs were frequently improperly imaged. Overcoming the illumination challenge was one of the main drivers for the development of new acquisition geometries. As a result of the significant effort undertaken by the acquisition contractor companies and oil companies, new acquisition geometries were developed to address this illumination challenge. WAZ and FAZ towed-streamer acquisitions were introduced in 2005 and 2010, respectively. Figure 1 shows the acquisition configuration and the offset azimuth distributions for NAZ, WAZ (Regone, 2007), and FAZ (circular shooting) towed-streamer acquisitions (Moldoveanu and Kapoor, 2009). The typical maximum offset for NAZ acquisition is a function of the cable length and is typically between 8000 and 14,000 m. As there are various technical limits to the lengths of streamers that can be towed, a solution to increase the maximum offset for NAZ acquisition was to add a source vessel in front of the streamer vessel (Figure 2). The maximum offset for WAZ acquisition is between 9000 and 10,000 m, and for FAZ multivessel circular shooting, which we refer to as dual vessel circular shooting geometry acquisition, it is between 14,500 and 16,500 m.

A parallel effort was dedicated to ocean-bottom seismic acquisitions implemented first with ocean-bottom cables and subsequently with OBNs. An OBN-type acquisition enables the practical implementation of acquisition geometries with FAZ and long to ultralong offsets to achieve the target imaging and velocity model building requirements (Figure 3). In towed-streamer acquisition, the maximum offset may be limited by the equipment, for example, the streamer length, number of streamers, and source vessels used. However, for OBN acquisition, the logistical limitations regarding towed streamers do not apply. OBN acquisition was previously used mainly for 4D reservoir development studies due to the important benefits in terms of the repeatability of the source and receiver locations, improved S/N, and improved frequency content, particularly at low frequencies. In addition, OBN surveys were used for reservoir development in congested fields or in areas where the S-wave recordings were useful for imaging through gas clouds. A typical sampling of the receivers for development studies is between 200 and 500 m, or denser, and shot sampling is between 25 and 100 m.

A critical step forward was made recently in seismic exploration by the introduction of OBN for very large exploration programs using sparse OBN-type acquisition geometries. By increasing the receiver sampling from a few hundred-meter ranges to a much sparser distribution of 1000–1200 m, seismic data with FAZ, ultralong offset, and good S/N at very low frequency can be acquired at an affordable cost over a large area. These developments have been at least partially enabled by the extended battery life that allows nodes to be deployed for a much longer period on the seabed.

Seismic modeling was performed to demonstrate that sparse OBN geometry could image accurately the subsalt reservoirs if the velocity model is accurate (Vigh et al., 2021a). Based on the modeling results and the previous experience gained in velocity model building using FWI and OBN data from previous development projects, the first sparse OBN survey was acquired successfully in 2019–2020 in the GOM, Mississippi Canyon area. To increase productivity, a simultaneous source was used by using two source vessels with a triple source deployed on each vessel.

The typical seismic source that has been used widely in marine seismic acquisition is the airgun source. The FWI requirement to have a very good S/N at very low frequencies, 1–2 Hz, pushed the seismic industry to develop new low-frequency seismic sources. Recently, several new types of low-frequency seismic sources were developed. One example is the WolfSpar, a vibratory type of source, which can generate seismic data with good S/N at large offsets in the frequency range of 1.7–2.5 Hz. This source was successfully tested in the GOM for OBN acquisition in 2018 and FWI produced a very high-quality velocity model (Brenders et al., 2020).

OBN seismic data with ultralong offsets require a large patch of source positions surrounding the patch of receiver (Figure 3). The shot halo around the receiver patch is designed to achieve the maximum desired offset. To be able to economically acquire a sparse OBN survey for a large exploration program, it is critical to use a static receiver patch. Acquiring shots in a minimum time is only possible with simultaneous shooting (Bansal et al., 2013) using multiple source arrays on each source vessel. Hence, simultaneous shooting sparse node OBN acquisition is quickly evolving as a new standard for seismic surveys in complex geologic areas such as deepwater GOM.

As we will show in the paper, seismic data from such sparse OBN surveys using low-frequency sources and simultaneous shooting were able to produce an accurate velocity model, particularly in the subsalt area due to FAZ coverage, ultralong-offset penetration, and good S/N at very low frequencies.

The power of FWI is determined by three key factors in the data collection: first is the low-frequency content of the observed data, second is the maximum offset length, and third is the azimuthal distribution of the offsets. In complex geologic environments such as salt provinces, the initial model is another crucial part of the FWI success. The background trends need to be correctly introduced, especially for the long offsets to avoid the possible large cycle skipping of the diving waves. Meanwhile, the gross salt geometry errors need to be revisited before running FWI, especially when the offset is limited rather than 50+ km.

We demonstrate for all major acquisition techniques such as the WAZ, FAZ, and the sparse OBN with shorter versus ultralong offset from limited azimuthal to full azimuthal data collection that FWI can successfully produce highly detailed velocity models. The seismic sources used in all cases are air gun and some with finetuning of the sources enable the recording of frequencies as low as 1.6–1.8 Hz.

In the WAZ and FAZ or similar long-offset streamer-data acquisition in a deepwater environment, the reflections dominate the recording. The data lack deep penetrating diving waves and therefore can only be used to modify mid- and small-wavelength anomalies rather than long-wavelength errors. Even if the water depth is shallow, due to limited diving wave penetration, only the reflection energy can help in updating the deep part of the model. Nevertheless, whether it is a shorter offset or an ultralong offset, FWI uses the full wavefield for updating the velocity model. Although the short-offset updates are dominated by reflection energy, the long offset mostly brings out the diving wave energy which will dominate at least the larger early updates.

In Figure 4, we compare the full-record FWI kernels at low and high frequencies using a source-to-receiver distance representing different maximum offsets that are reflected to the various acquisition geometries. It is important to see the differences between the acquisition geometries to understand what FWI can achieve in complex geology. For this exercise, the BP 2004 salt model was used to demonstrate the FWI kernel difference with different offsets and frequencies. The first observation we can make is that the low-frequency kernel produces low-wavenumber updates, especially along the diving wave path indicated by the yellow arrows in Figure 4. The simpler gradient resulting from low-frequency data is particularly helpful in making large kinematic model updates, such as reshaping the salt. Second, the long-offset length obviously gives deeper penetration that allows kinematic updates at the subsalt and deeper region, which is our target with FWI. The next step is to see how the full wavefield illuminates the given geology differently at the deep target, which is an essential part of the survey design. Figure 5 shows how this study can be created with prior knowledge of the geology. Using a detailed geologic area to create velocity and density fields, the velocity is perturbed by a small amount, e.g., 1%, creating synthetic data with given shooting geometry whether it is NAZ, WAZ, circular shooting, or sparse node OBN. We run an iteration of FWI using the original 100% models to produce an updating direction proportional to the perturbation. This volume will represent the illumination of the chosen acquisition of the given geologic setting.

Figure 6 shows the different diving wave illuminations using different offset lengths which relate to acquisition differences and highlights why the current OBN long-offset acquisition is superior to the previous acquisition techniques to achieve a well-desired deep kinematic update in the velocity model. This could be a very important step prior to an acquisition design with a given geology. The FWI response is different in different basins around the world so there is no single solution. In this paper, we demonstrate most of the findings in the GOM. We notice that this is not an example applicable to every geology. A specific FWI experiment needs to be carefully assessed for different geologic objectives.

In the past few decades, predominantly WAZ data were acquired for velocity model building and imaging. Therefore, we start with the WAZ data acquisition to show how FWI can perform. If we rely on diving waves only, then the updates are not going to penetrate to deeper depths. Most updates, depending on the velocity gradient of the initial model, will be up to the depth of one-third of the maximum offset. When extending offset lengths from 8 to 16 km, similar to the case of dual vessel circular shooting geometry acquisition, the updating power of FWI still cannot penetrate deep enough for some deepwater settings to reach the subsalt areas. Although the longer offset helps to record a more complete wavefield, it is not sufficient in the complex salt environment to record the most valuable diving waves going through and below the salt bodies. The 16 km offset allows an approximate maximum of 5–6 km diving wave penetration depending on the water depth. This may allow FWI to refine the top-of-salt shapes but barely gives a chance to change the base of salt and will not detect the salt feeders either. That is the reason the deep model needs to be fairly accurate. Otherwise, different models still produce similar misfits in the FWI that gives high uncertainty to the model building process, especially for the deep part of the model, when using towed-streamer data sets, which only record a limited wavefield. Because of the aforementioned limitation, often we must resort to using the full record to achieve deeper updates and use low frequencies to bring kinematic changes to the image via the velocity model updates. Sparse node OBN acquisition overcomes penetration limitation by introducing ultralong offset. With up to 50 km offset, the wavefield in the observation can be captured more completely. However, FWI still benefits from using the full shot record for a velocity update even though the sparse node OBN data are diving wave rich.

We will demonstrate via two of the extreme acquisitions, the orthogonal WAZ and the sparse node with short versus ultralong offset, how FWI can successfully produce highly detailed velocity models. In terms of azimuthal coverage, the orthogonal WAZ can give a semifull azimuthal coverage up to 8 km offset due to the two-directional WAZ shooting, whereas the nodal acquisition is by default full azimuthal all the way up to the designed nominal offset of approximately 50 km. The seismic sources are in both cases air guns. Even with finetuning of these sources, approximately 1.8 Hz is the lowest frequency that was injected during the data acquisition.

The WAZ data acquisition with an 8 km maximum offset in a deepwater environment is deficient in diving waves but rich in reflections, and therefore they can only be used to modify mid- and small-scale wavelength errors rather than long-wavelength inaccuracies. The low-wavenumber update may be derived from the reflection backscatter energy which can be derived implicitly with the energy norm imaging condition or explicitly with Born-based modeling with the full-wavefield modeling or only focusing on the classical full-wavefield modeling with low frequencies at approximately 3 Hz. This approach sometimes can produce misleading updates due to the small angle of incidence in the data, which is a warning sign that the updates could be nongeologic, even though it produces flat image gathers. In our WAZ example, thanks to a larger angle of incidence, we take advantage of the low-wavenumber update coming from the reflection-based FWI (Sun et al., 2016; Vigh et al., 2016), despite the relatively good starting model, to make medium-scale changes in the model, especially in the deep part of the velocity model before switching to the full-record FWI to address the small-scale anomalies in the model. Figure 7 shows the results of the FWI refinements that can be obtained by short-offset data. Figure 7a shows the legacy image and Figure 7b shows the FWI-updated velocity and enhanced image quality, especially in the deep section, particularly at the top of the Cretaceous, which is an excellent control for the velocity accuracy above.

Increasing the offset length to 16 km maximum offset and azimuthal coverage to full 360° in the acquisition results in an elevated resolving power of FWI in depth penetration and salt geometry reshaping. Figure 8 demonstrates that FWI achieved significant improvements in the subsalt sedimentary delineation and flattened or simplified the base Louann salt (Figure 8b) compared with the legacy model and image (Figure 8a). In the towed-streamer examples (Figures 7 and 8), the initial model of FWI is derived from the legacy model by smoothing the model according to the starting frequency in the FWI iterations.

OBN acquisition, especially when conducted for complex model building with FWI, allows the survey design community to relax certain design parameters for the surveys, especially the node sampling, as with sufficient shot coverage the data can be input to FWI in the common-receiver domain. This significantly reduces the overall acquisition cost but maintains the ultralong-offset length beyond 50 km. In the field data example that we are considering, the sparse node acquisition encompasses a 1.2 km distance between the nodes with a 50 m × 100 m shot carpet and a 20 km halo around the nodes. The natural bin size is 25 m × 50 m. The data were acquired by two vessels using simultaneously fired triple sources, with a 1000 ms active dither, and using continuous recording.

We must assess the maximum acceptable node spacing to perform FWI reliably at frequencies up to 20 Hz, which is the upper limit in the FWI model building that can result in meaningful kinematic change in imaging, although higher frequencies FWI may be desired to derive a high-resolution pseudoreflectivity, sometimes called the full-wavefield image (Sheng et al., 2022; Mcleman et al., 2023; Wei et al., 2023).

The ultimate goal of FWI is to include higher-fidelity physics in the forward modeling to closely mimic the wavefield propagation in the real earth to achieve a more accurate velocity update. The EFWI is naturally the next step in the industry especially when the geologic challenges are present, such as in the GOM where the salt geometries are complex and recovering the subsalt sedimentary sections is difficult.

Figure 13 shows clear differences between the elastic and acoustic behavior especially beyond the critical angle where the elastic shot record has a phase rotation compared with the acoustic one shown in Figure 13a versus 13b. This indicates that it will be challenging to correctly update the top of salt or the entire salt geometry itself by the acoustic FWI. When the frequencies are increased the differences become much less obvious, though the amplitude treatment is better with elastic propagation.

Here, we reveal an EFWI example to demonstrate that more accurate physics makes a difference in complex geology. The initial model was derived from a legacy model that was built with the traditional top-down approach. We first smoothed the legacy model and performed rigorous QCs to validate the initial model. FWI was then used to update the entire model, including the salt geometry and sedimentary sections shallow and at depth, in a multiscale manner, starting from 1.7 Hz. The frequency-marching increment was 1 Hz at the low end to explore the power of low-frequency signals to avoid cycle-skipping problems. We executed acoustic and elastic FWI in parallel to determine whether we can see an improvement in the model update by using more accurate physics using real field data examples (Shen et al., 2020; Plessix and Krupovnickas, 2021). The low-frequency updates modified the salt and the subsalt velocities to achieve better kinematics that improve the deep velocity field up to the base Louann Salt as well as images for subsalt regions in acoustic and elastic FWI. Some obvious conclusions can be drawn after finishing the low-frequency updates. Figure 14d shows that the EFWI results have more geologic consistency compared with the acoustic result in Figure 14a. The update in the top of salt from EFWI (Figure 14e) is slowing down instead of speeding up as shown in the acoustic FWI update (Figure 14b). The slowing down in the top of salt sediment layer is in fact the right direction due to the smoothing in the initial model that tends to increase the top of salt sediment velocity. The subsalt updates are more consistent in the EFWI update, and this can be attributed to the better amplitude handling in EFWI in terms of the residual fields. Finally, the steeply dipping events in the image from the EFWI-updated model (Figure 14f) are more evident than those in the comparison image using the acoustic FWI-updated model (Figure 14c), which suggests that the EFWI could generate more realistic model updates than acoustic FWI (Vigh et al., 2022). The EFWI provides better phase alignment between the observed and the simulated data. This is one of the reasons why kinematics are better at high-velocity interfaces, especially at low-frequency inversion. The amplitude fidelity is also better defined in the EFWI result. This is why the results show better conformance to the geology and improve the image below the salt bodies too.

These advantages of EFWI will shine in complex geologies wherein high-fidelity physics requires obtaining geologically plausible results.

In the past few decades, FWI has progressed to show that data-driven model building is a game changer to improve model quality. The objective function improvement and the acquisition evolution resulted in increasing justification that FWI is the way to update or build earth models, thus confirming the projections from the early pioneers of FWI in the 1980s. Increasing the offset length, providing full azimuthal coverage, and pushing the low frequencies down below 2 Hz enabled and made clear FWI is the tool to solve development and exploration challenges in complex geologic settings. Despite the early elastic derivation, we have executed acoustic FWI due to the cost implication for more than a decade. Thanks to rapid advances in computer power, EFWI is emerging on the horizon as a practical solution. Introducing the EFWI, one can assume that the technique allows us to account for and obtain more of the physical parameters of the rock types of the subsurface due to a more realistic description of the seismic wave propagation and more physical attributes involved in the elastic wave equation. However, the challenge remains as to how to use EFWI to input all components into the inversion to gain the full power of FWI with given multicomponent acquisitions.

The authors thank SLB EXD Multiclient and TGS for permission to show and publish this work.

Data associated with this research are confidential and cannot be released.

Biographies and photographs of the authors are not available.