A method to impose sparsity on full-waveform inversion and least squares migration

The method improves seismic inversion by applying a sparsity-promoting norm and Hessian approximation to refine seismic models, addressing convergence and noise issues, resulting in accurate subsurface property estimation.

WO2026236586A1PCT designated stage Publication Date: 2026-11-12DOWNUNDER GEOSOLUTIONS (AMERICA) LLC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027230
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-05-08
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing seismic inversion methods struggle with convergence issues, spurious solutions, and noise introduction due to the ill-conditioned nature of the Hessian matrix, leading to inaccurate estimation of Earth's subsurface formation properties.

Method used

A method involving full-waveform inversion and least squares migration with a sparsity-promoting norm, using a quadratic constraint and Hessian approximation, to refine seismic models and reduce computational expense.

Benefits of technology

This approach enhances the accuracy of subsurface property estimation by reducing spurious oscillations and noise, while maintaining data fit, thus providing a more precise spatial distribution of Earth formation properties.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027230_12112026_PF_FP_ABST
    Figure US2026027230_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A method for estimating the spatial distribution of Erath formation properties in Earth's subsurface from measurements of seismic signals made proximate a volume of the subsurface to be evaluated, includes generating an initial model of the spatial distribution. The initial model and measured seismic signals are entered into an inversion process to obtain an updated model of the spatial distribution using the measured seismic signals. A Hessian is estimated using data obtained during the inversion step that accepts the initial model as input to generate an updated model. A transform of the updated model is determined. A secondary inversion on the transformed updated model is performed, with a sparsity promoting norm with a constraint being an approximation of an objective function used in the initial inversion to at least second (quadratic) order to obtain a new transformed updated model that will ensure the quality of the data-domain fit is within a specified tolerance. The new transformed updated model is inverse transformed. An output of the inverse transform is generated as an image of the spatial distribution.
Need to check novelty before this filing date? Find Prior Art

Description

PATENT APPLICATION ATTORNEY DOCKET NO. DGL-25-01PCTA METHOD TO IMPOSE SPARSITY ON FULL-WAVEFORM INVERSION AND LEAST SQUARES MIGRATION BACKGROUND

[0001] This disclosure relates to the field of seismic exploration of the Earth’s subsurface.More particularly, the disclosure relates to inversion processing of seismic data to determine the most likely spatial distribution of Earth subsurface formation parameters related to propagation of acoustic or elastic waves through such formations.

[0002] Many challenges in the oil and gas industry involve determining the Earth’s subsurface formation properties from data obtained by seismic acquisition. One or more seismic energy sources are used to impart seismic energy into the Earth, which then propagates as a seismic wavefield and interacts with elastic heterogeneities. As a result, some of the seismic energy returns to one or more receivers disposed on the land surface, water bottom or water surface, which measure properties of the wavefield as a function of time and location. The source(s) can include, but are not limited to, compressed air guns, vibratory transducers, explosives, or natural sources. The seismic wavefield can contain a superposition of different wave types, such as reflections, refractions, and converted (e.g., compressional to shear) waves. Combinations of these types of waves are used in many different processing techniques, such as inversion, to help estimate the Earth’s subsurface properties and their spatial distribution. Examples of these estimated properties, which may be referred to as “model” properties or “parameter classes” include velocity (wave speed), seismic anisotropy, acoustic impedance, amplitude versus offset / angle (AVO / A) parameters, reflectivity, and the spatial distribution of any or all of the foregoing parameters.

[0003] Due to the nature of the Earth and of seismic signal acquisition, it is impracticable to simply obtain a unique and correct estimate of the model properties directly from the seismic data. There can be many sets of model properties (“solutions”) that can satisfy the observed (measured) seismic data in some sense. Many of these solutions will not be geologically plausible. While it is impossible in these circumstances to directly obtain thePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTtrue solution, an estimate of the true solution can be obtained through a process known as inversion which uses assessments of how well the model properties fit with the observed seismic data. The process of inversion uses a forward modeling operator, which synthesizes the seismic wavefield that would be observed if the Earth’s subsurface had been correctly described by the model properties. The forward modeling operator may be some appropriate form of the wave equation. The resulting synthetic (“forward modeled”) seismic data are then compared with the recorded (“observed”) seismic data and any differences between the two are assumed to be due to errors in the estimate of the model properties. The difference between the forward modeled and observed seismic data is termed the “residual” vector. The amount of residual can be quantified by a “cost” or “objective” function, which, for example, could be given by the square of its “L2 norm”. The goal of inversion processing is to find model properties which minimize this objective function. In such case, the modeled data match the recorded or observed data in a leastsquares sense.

[0004] If the model properties are not optimal for a particular forward modeling operation, then it is necessary to change the model parameters in a way that improves the objective function. This may be obtained by applying an adjoint of the forward modeling operator (sometimes referred to as a “migration” operator) to the residual vector, and applying an imaging condition which determines the gradient of the objective function with respect to the model parameters. The gradient is the direction in which the model needs to change to improve the match of the modeled data to the observed data. The set of new or updated model properties that will improve the fit or match may be calculated by an optimization algorithm such as steepest descent or the limited-memory Broyden-Fletcher-Goldfarb- Shanno algorithm (L-BFGS). The inversion process steps are repeated until the objective function is sufficiently small or until some other stopping criterion is satisfied. Formally, an example of an objective function is given by the square of the L2 norm, / (m) = | ||d - L(m)||^ (1)PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT

[0005] where J is the objective function, L is the forward modelling operator which describes how to synthesize data from the model properties, m, and d is the observed or measured data.

[0006] There are many different optimization algorithms that can be used to solve inversion problems, each with its own strengths and weaknesses. For example, first order optimization methods such as steepest descent only make use of the gradient and the model parameters from the previous iteration. For a single iteration, this is simple and computationally inexpensive. However, because it suffers from parameter class cross-talk and parameter class scale differences, convergence to a final model can be slow and may require many iterations. The convergence rate of steepest descent can be improved by exploiting a more extensive history of previous steps, rather than simply relying on the most recent prior step. Examples of this are “momentum” techniques, “adaptive gradient” approaches (“AdaGrad”, Duchi et al., 2011). However, for complex non-linear problems, higher-order optimization schemes tend to perform better.

[0007] Second-order optimization approaches make use of the gradient (first derivative of the objective function with respect to the model parameters) and second derivative information. The second derivative is described by the Hessian matrix (sometimes simply referred to as the “Hessian”), which succinctly provides information about the curvature of the objective function. “Newton’s method” (Newton, 1711; Raphson, 1690) is an example of a second-order optimization scheme using the Hessian. As a result, it converges much faster than steepest descent and is much more resilient to cross-talk between parameter classes and to scaling issues. However, implementation of Newton’s method on large problems with billions of unknowns is not practical because it requires that the Hessian matrix is formed explicitly and that the inverse Hessian is well defined. Consequently, approaches that approximate the inverse Hessian matrix at each iteration, known as “quasiNewton” methods, have been developed (e.g., BFGS). BFGS uses the complete history of all iterations to build an approximate inverse Hessian matrix. However, for large inversion problems, even using this approach, it is prohibitive to store the necessary information from all previous iterations. A modification of this approach, L-BFGS, uses a low-rank approximation of the inverse Hessian matrix using only a few of the most recent iterations,PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTinstead of the complete history. L-BFGS has been shown to be practical for large-scale inversion problems and typically converges faster than steepest descent.

[0008] The sensitivity of the objective function to the model parameter classes varies for several reasons, including the geometry of acquisition (the arrangement of sources and receivers) and the geologic structure of the Earth. Note that “sensitivity of the objective function to the model parameter classes” here refers to how much a change in the model parameter class affects the forward modeled seismic data. During seismic data acquisition, unless all parts of the sub-surface in the survey area were probed equally by the seismic waves there may be weak or no information about some regions of the Earth in the recorded data. As a result, convergence will be slow and the resulting model updates can be spurious in these locations. If the full Hessian could be computed and it was sufficiently well conditioned to be inverted, then the inverse Hessian would better compensate for illumination and cross-talk too. Unfortunately, lesser approximations used in quasiNewton methods have a more limited ability to obtain illumination compensation and resolve cross-talk, and the Hessian itself can be poorly conditioned. As a result, it is common practice to additionally include “preconditioners” to assist in illumination compensation to help accelerate convergence in quasi-Newtonian inversion schemes.

[0009] In the regions of the subsurface that are poorly constrained by the data, it is possible for there to exist a set of solutions all of which give approximately the same data domain (and objective function) response, or have no effect on the data domain at all. Such solutions are said to sit in the “null-space” and can be a consequence of the underdetermined nature of some locations in the model. Therefore, it is difficult for the inversion to distinguish which of these solutions represents the true earth geology (spatial distribution of formation properties) without including regularization in the inversion, and so the solutions generated can contain implausible earth structures and noise. Methods to understand this uncertainty and the sensitivity of different locations in the model and their effect in the data domain can be obtained via a technique called “null-space shuttling” (Deal and Nolet, 1996; Fichtner and Zunino, 2019; Keating and Innanen, 2020; Keating and Innanen, 2024). The use of “preconditioners” can help to guide the inversion towards more desirable solutions in the absence of suitable data constraints, such as structurallyPATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPToriented smoothing (Fehmers and Hocker, 2003), or by imposing additional weighted penalty terms to the objective function, such as Tikhonov regularization, or a weighted sparsity promoting LI constraint on the inverted models (or transformed domain versions thereof). However, the parameterization of such preconditioners often requires careful and extensive testing to ensure the correct balance between appropriately conditioning the inversion but not so much that the genuine information can become biased or lost. As a result, parameterizing preconditioners can be non-trivial and expensive to test when the forward modeling operator L is computationally intensive to execute. Determining the appropriate weighting factor for the additional penalty terms on the objective function also suffers from the same parameterization issue, but such sparsity promoting penalty terms can have little effect on the final result if a small number of iterations are used to determine the updated model parameter classes, again due to the computational expense of evaluating the operator L.

[0010] Acquired seismic data often contains noise which can lead the inversion towards solutions that are undesired. Even if the noise is weak in the data domain, again due to the ill-conditioned nature of the Hessian it is possible for noise with large amplitudes to be introduced into the final model result to match the noise in the data domain. This is sometimes referred to as data “over-fitting” and is strongly related to the “null-space” issue. This is where large changes in the amplitude of the inverted models along some step directions correspond very weakly to some directions in the dataspace. Thus, it results in large model amplitudes to fit the noise in the dataspace, which is undesirable.SUMMARY

[0011] A method according to one aspect of the present disclosure is for estimating a spatial distribution of Earth formation seismic properties in the Earth’s subsurface from measurements of seismic signals made proximate to a subsurface volume. A method according to this aspect includes generating an initial model of the spatial distribution. The initial model(s) is entered into an inversion process to obtain an updated model of the spatial distribution using the measured seismic signals in the inversion process. A Hessian of the updated model is estimated. A transform of the updated model(s) is determined. APATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTsecondary inversion on the transformed updated model(s) is performed, with a sparsity promoting norm with a constraint being an approximation of an objective function used in the initial inversion to at least second (quadratic) order to obtain a new transformed updated model(s) that will ensure that the quality of the data-domain fit is within a specified tolerance. The new transformed updated model(s) is inverse transformed. An output of the inverse transform provides an image of the spatial distribution of the seismic properties.

[0012] A non-transitory computer readable medium according to another aspect of this disclosure has stored thereon a computer program having logic operable to cause a programmable computer to perform the above-described method.

[0013] In some implementations, inversion process comprises at least one of full waveform inversion or least squares migration.

[0014] Some implementations further comprise determining a Hessian approximation using a combination of data obtained at previous iterations steps from the inversion process when obtaining the updated model from the initial model.

[0015] In some implementations, the sparsity is imposed in a transformed domain.

[0016] In some implementations, the transformed domain comprises at least one of curvelet, Fourier, Radon, or wavelet domains.

[0017] In some implementations, the sparsity with quadratic constraint is imposed in overlapping spatial windows.

[0018] In some implementations, the Hessian approximation is regularized and / or constrained.

[0019] In some implementations, the Hessian approximation is constrained to be positive definite.

[0020] In some implementations, the Hessian approximation is determined using stabilized non- stationary convolution operators.

[0021] In some implementations, a basis pursuit approach is applied at or after one or more of the iterations of the inversion process.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT

[0022] In some implementations, at least one sparsity imposed model is determined by thresholding in a transformed domain, which is validated with the quadratic constraint.

[0023] In some implementations, a priori information is used to constrain the additional inversion.

[0024] In some implementations, the new updated model is determined using iterative thresholding in a transformed domain.

[0025] In some implementations, the thresholding is based on the amplitude, energy, and / or phase of the transformed models.

[0026] In some implementations, the models and / or gradients are preconditioned.

[0027] In some implementations, the sparsity is imposed using a sparsity promoting norm comprising, LP-norm, Nuclear-norm, total variation norms, or combinations thereof.

[0028] In some implementations, the a priori information comprises a subsurface structural dip.

[0029] Other aspects and possible advantages will be apparent from the description and claims following.BRIEF DESCRIPTION OF THE DRAWINGS

[0030] FIG. 1 shows an example arrangement of seismic sources and seismic sensors that may be used to acquire seismic data for processing in accordance with the present disclosure.

[0031] FIG. 2 shows a flow chart of an example implementation of a method according to the present disclosure.

[0032] FIGS. 3A, 3B and 3C show, respectively, the Ricker wavelet source, the true earth reflectivity for this synthetic, and the synthetic observed “field” data.

[0033] FIGS. 4A and 4B show, respectively estimated reflectivity from least-squares inversion and the corresponding modeled data overlain with the observed data.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT

[0034] FIGS. 4C and 4D show, respectively, estimated reflectivity from constrained basis pursuit and least-squares estimated reflectivity, and the observed data overlain with the modelled data using the constrained basis pursuit sparse reflectivity.

[0035] FIGS. 5A, 5B and 5C show, respectively, true reflectivity, LS-RTM recovered model, and sparse reflectivity derived using the constrained basis-pursuit algorithm with the LI -norm with quadratic constraint results in a sparser reflectivity model without compromising the data fit to within a specified tolerance.

[0036] FIGS. 6A, 6B and 6C show, respectively, the synthetic true-earth P-wave velocity model, the velocity model inverted with FWI using the least-squares based objective function, and the basis pursuit result takes the FWI velocity as input and performs basispursuit with a total variation objective function with the quadratic constraint thereby promoting sparsity (“blockyness”) in the model. The basis-pursuit result reduces spurious oscillations in the velocity model without compromising the data fit to within a specified tolerance.

[0037] FIG. 7 shows a comparison of true and inverted velocity model profiles. The basis pursuit result promotes sparse discontinuities and reduces small oscillations in the model.

[0038] FIG. 8 shows an example implementation of a computer system that may be used to perform a method according to the present disclosure.DETAILED DESCRIPTION

[0039] FIG. 1 shows a vertical section view of an ocean bottom cable (OBC) seismic survey being conducted using, for example, two different “source” vessels for towing seismic energy sources. In some embodiments, only one source vessel may be used. According to the present disclosure, any number of source vessels may be used and the following description is not intended to limit the scope of the present disclosure. The source vessels move along the surface 16A of a body of water 16 such as a lake or the ocean. In the present example, a vessel referred to as a “primary source vessel” 10 may include equipment, shown generally at 14, that comprises components or subsystems (none of which is shown separately) for navigation of the primary source vessel 10, for actuationPATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTof seismic energy sources and for retrieving and processing seismic signal recordings. The primary source vessel 10 is shown towing two, spaced apart seismic energy sources 18, 18 A.

[0040] The equipment 14 on the primary source vessel 10 may be in signal communication with corresponding equipment 13 (including similar components to the equipment on the primary source vessel 10) disposed on a vessel referred to as a “secondary source vessel” 12. The secondary source vessel 12 in the present example also tows spaced apart seismic energy sources 20, 20A near the water surface 16A. In the present example, the equipment 14 on the primary source vessel 10 may, for example, send a control signal to the corresponding equipment 13 on the secondary source vessel 12, such as by radio telemetry, to indicate the time of actuation (firing) of each of the sources 18, 18A towed by the primary source vessel 10. The corresponding equipment 13 may, in response to such signal, actuate the seismic energy sources 20, 20A towed by the secondary source vessel 12.

[0041] The seismic energy sources 18, 18A, 20, 20A may be air guns, water guns, marine vibrators, or arrays of such devices. The seismic energy sources are shown as discrete devices in FIG. 3 to illustrate the general principle of seismic signal acquisition. The type of and number of seismic energy sources that can be used in any embodiment is not intended to limit the scope of the disclosure.

[0042] In FIG. 31 an OBC 22 is deployed on the bottom 16B of the water 16 such that spaced apart seismic receiver modules 24 are disposed on the water bottom 16B in a preselected pattern. The receiver modules 24 may include a pressure or pressure time gradient responsive seismic sensor, and one or more seismic particle motion sensors, for example, one-component or three-component geophones, or one- or three-component accelerometers (none of the sensors are shown separately). The type of and the number of seismic sensors in each module 24 is not intended to limit the scope of the disclosure. The seismic sensors in each module 24 generate electrical and / or optical signals (depending on the sensor type) in response to, in particular, detected seismic energy resulting from actuations of the seismic energy sources 18, 18A, 20, 20A. In some embodiments, thePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTsensors may comprise pressure or pressure time derivative sensors such as hydrophones, and one or more particle motion responsive sensors such as geophones or accelerometers. In some embodiments, such particle motion responsive sensors may be oriented so as to be sensitive principally (ignoring any cross-component coupling for purposes of explanation) to vertical particle motion. The signals generated by the various sensors may be conducted to a device near the water surface 16A such as a recording buoy 23, which may include a data recorder (not shown separately) for storing the signals for later retrieval and processing by the equipment 14 on the primary source vessel 10, or other processing equipment to be described further below. The data storage functions performed by the recording buoy 23 may be performed by different types of equipment, such as a data storage unit on a recording vessel (not shown) or a recording module (not shown) deployed on the water bottom 16B, e.g., proximate each sensor module 24, or even on the primary or secondary source vessels. Accordingly, the disclosure is not limited in scope to use with a recording buoy or any other specific recording device(s).

[0043] Although the description of acquiring signals explained with reference to FIG. 1 is for sensors deployed on the water bottom, it will be appreciated that it is possible to obtain corresponding measurements at any selected depth in the water, using, for example, seismic sensors disposed in a towed cable such as described in U.S. Patent No. 7,239,577 issued to Tenghamn, et al. Further, the present disclosure is not limited to use with seismic signals acquired in or on the surface of a body of water; methods described herein may also be used with seismic signals acquired on land or acquired with sensors situated within a production well or borehole.

[0044] In a method according to the present disclosure, an updated estimate of the subsurface model parameter class(es) may be obtained using, e.g., full-waveform inversion (FWI) (Tarantola, 1984), least-squares migration (LSM) (e.g., least-squares reverse-time migration (LS-RTM) (Duv eneck et al., 2021) or least-squares Kirchhoff depth migration (LS-KDM) (Nemeth and Schuster, 1999)), whereby at each FWI or LSM iteration, the estimate of the updated model(s), the computed gradient(s), and the objective function values are stored.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT

[0045] Then the foregoing information is used in a secondary iterative inversion process to determine a new model or set of models (m*) using a basis pursuit type approach, which minimizes a “sparsity promoting” cost function G. G could be, e.g., the L0 norm, the LI norm, LP-norm or the total variation (TV) norm given by, for example, the expression:G= ||T(7n*)| |1(2) subject to the constraint of / (m*) < (yj(m) (3)

[0046] The operator T is a domain transform operator that could be, for example and without limitation, an identity matrix or a transform such as a curvelet, Radon, or Fourier transform. The output of T may comprise a different set of basis coefficients. However, as explained previously, computing the objective function of the new model set J(m*) can be very computationally expensive when using, e.g., FWI or LSM. The computational expense is a result of the fact that executing the forward modeling operator L is computationally intensive, as it represents solving a form of the wave equation. It is possible to avoid such computational expense when evaluating the constraint of Eq. (3) by approximating the calculation of the objective function of the new model or model set J(m*) (which could be e.g., a least-squares objective function) to a second order (quadratic) approximation by the expression: / (m) + gT(m* — m) + |(m* — — m) < (4)where g is the gradient output from the inversion, e.g., FWI or LSM, m is the updated model from FWI or LSM, and H is the Hessian. Examples of how to estimate H are provided below.

[0047] a is a free parameter that can be used to control the extent to which the constraint will be allowed to be violated or enforced. Note that by using the quadratic approximation of the constraint, as in Eq. (4), it becomes no longer necessary to evaluate the forward modeling operator / , when applying the constraint, and thus the process is spared from thePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTcomputational expense of evaluating the forward modeling operator L. The new model set m* can be determined iteratively using the above equations with a constrained optimization method such as, but not limited to, Lagrange multipliers or trust-region methods.

[0048] An example of a first choice is to set o = 1; this makes the constraint enforce the fact that any » / * found must, if entered into the forward model L, satisfy the original data fit objective function at least as well as the original FWI or LSM result while imposing sparsity to the transformed model(s), T(ni).

[0049] An estimate of H can be determined in a number of different ways. For example, the gradients and models produced from each iteration of FWI or LSM can be input into a BFGS style approach (Nocedal and Wright, 2006), or it may be considered to use only the final few (or more) gradients and models to design non- stationary convolutional filters to approximate the Hessian / / using the expression:Fs ~ y (5) where 5 is the change in the model, y is the change in the gradient, and is a non- stationary convolutional operator (a “spatially varying matching filter”) approximating the Hessian. Note that stabilization methods may be used to ensure robust estimates of / / , for example, ensuring H satisfies the “positive definite” property and is symmetric. Such approaches to determine H and the new model m* can be performed using overlapping windows in ID, 2D, 3D, or even higher dimensions.

[0050] Note that while G in Eq. (2) as shown above makes use of a form of the LI -norm, in practice, any sparsity promoting norm combined with T, such as L0, LP or the totalvariation (TV) norm could be used. The TV-norm promotes sparsity in the spatial derivatives of the models. Note that for any sparsity promoting norm chosen (or combinations of such norms) they can act on each parameter class individually or jointly with several models simultaneously.

[0051] Depending on the size of the models m, and the number of iterations required to determine the new “sparse” models m* that satisfy the quadratic constraint in Eq. (4) using the basis pursuit approach, the computational expense can still be larger than desirable. ItPATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTis possible to simplify the problem further. Instead of solving the foregoing equations using a traditional optimization method, it is possible to implement a thresholding scheme such that the coefficients that form the basis of T(m) that do not satisfy the threshold are weighted downwardly. The resulting model set m* is checked to determine if it satisfies the quadratic approximation of the data constraint. If it does not, then decreasing the threshold continues by including in each iteration more non-zero basis coefficients until the constraint of Eq. (4) is satisfied. At such point, the sparse model found that satisfies the quadratic constraint but has the fewest number of non-zero basis coefficients may be considered as the solution model set m*. An example of such a threshold is the sum of the squares of the model samples found contributing to each basis coefficient. The foregoing approach can find a sparse model m* in much fewer iterations than using conventional optimization methods, thus making it computationally inexpensive to determine. A high- level flowchart showing an example of this simplified approach is shown in FIG. 2. Note that the choice of which basis coefficients to keep in this approach doesn’t need to be a strictly “top down” approach, one could treat this as a form of optimization where we seek the smallest possible number of combinations of basis coefficients that still satisfies the quadratic constraint.[00521 Note that the decision over which basis coefficients to keep and which to exclude can be further aided by auxiliary information, such as the expected geological dip of the subsurface structure in the spatial window under consideration. Likewise, the value of cr could vary between spatial windows.

[0053] Each basis coefficient in the transformed model(s) T(m) in this case is effectively given a "score." Higher scores mean that such basis coefficients will likely contribute substantially to preserving the data domain match as quantified via the quadratic approximation of the objective function, e g., Eq. (4). Lower scores mean that such basis coefficients either contain noise that is desirable to remove, or contribute very little to preserving the data domain match. Therefore, it is possible to set a threshold initially which, for example, disregards (removes) all basis coefficients except for the highest score basis coefficient. The threshold is effectively just a number, which could be spatially varying and which can be selected as a supplied input to the process that is compared with the scorePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTfor each basis coefficient. This will provide, when an inverse transform is applied to the thresholded basis coefficients, a new model m * where the effect of this decision on the data domain (the domain of seismic signals) match can be checked using the quadratic approximation of the objective function as in Eq. (4). If the match is significantly worse, then the initial threshold is rejected and the threshold is changed to make it more lenient and allow retaining basis coefficients in the new model transform with lower scores to be included. This process repeats until the data domain match using the quadratic approximation of the objective function of Eq. (4) is satisfied to within a predetermined tolerance or until all basis coefficients are included. The "score" can be calculated in a number of different ways, for example, using a priori information, basing the threshold selection on the amplitude, energy, and / or phase of the transformed models, or where the a priori information contains details about subsurface structural dip of the Earth formations

[0054] An example implementation of a method according to the present disclosure may be better understood with reference to the flow chart in FIG. 2. At 30, an initial model of spatial distribution of one or more properties of subsurface Earth formations, and recordings of seismic signals made, e.g., as explained with reference to FIG. 1 are entered into a computer or computer system (explained further below with reference to FIG. 8). At 32, inversion such as full waveform inversion (FWI) or least squares migration (LSM) may be performed in the computer using the input initial model(s) and recorded signals to obtain an updated model m of the spatial distribution. At 34, a transform operator is applied to the model m to obtain a transformed model / ( / ??) with the new basis coefficients. At 36, a threshold is selected whereupon the thresholded transformed model is converted to an initial transformed model(s)At 38, any coefficients in the transformed model(s) ' / ( / ??) that do not meet the threshold are discarded, as explained above, to provide T(m*). At 40, an inverse transform 7 of T(m*) is performed after rejection of the below-threshold coefficients to provide an updated model(s) m*. Concurrently, at 44, an estimate of the Hessian H of the FWI / LSM model is determined and used as input to an evaluation at 42. At 42, the updated model(s) m* is evaluated to determine whether it meets a quadratic constraint with reference to the estimate of the Hessian H from 44. If not, the process returns to 36, wherein the threshold is decreased, and the process is once again performedPATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTas explained with reference to 38 and 40 in FIG. 2. The process may be repeated at 36, 38 and 40 until the quadratic constraint at 42 is satisfied. The process then ends at 46 with a final version of the updated “new” model m*. Examples of output from a method according to the present disclosure will be explained below with reference to FIGS. 3A through 7.

[0055] The updated “new” model(s) m* may be used in any known form (e.g., graphic visualization, numerical tables) as a representation and / or image of the spatial distribution of properties of Earth formations in the subsurface area being evaluated. Such representation may be used, for example and without limitation, in hydrocarbon exploration, subsurface reservoir monitoring, civil and / or geotechnical engineering.

[0056] It is important to note that the description of the method up to this point which has made use of the basis-pursuit style inversion to determine the sparsified updated models only made use of a constraint that was a quadratic approximation of the objective function used during the FWI or LSM stage, however, additional constraints can be included in addition to this. For example, to determine updated models with further geological conformability without compromising the data fit to within a specified tolerance, one could include additional constraints related to the rock-physics properties, such as facies-based rock physics constraints (Kamath et al., 2017; Singh et al., 2018; Zhang et al., 2018) or petrophysical constraints (Rocha and Sava 2018; Aragao and Sava, 2020). Note that such constraints are typically, but not always, guided by information obtained from well data, Vertical Seismic Profile (VSP) data, and / or borehole data. Such data can include, but is not limited to, sonic logs, checkshots, density logs, Vs / Vp logs, and gamma ray logs.

[0057] FIG. 3 A shows the Ricker wavelet source, FIG. 3B shows the true earth reflectivity for this synthetic, and FIG. 3C shows the synthetic observed “field” data. The wavelet and observed data are used as inputs to invert for an estimate of the reflectivity (FIGS. 4A and 4B). The corresponding modelled data associated with the inverted reflectivity is also shown in FIGS. 4B. It may be observed that despite the excellent match between the modelled and observed data, the derived reflectivity contains undesirable oscillations.

[0058] Estimated reflectivity from least-squares inversion are shown in FIG. 4A and FIG.4B shows the corresponding modelled data overlain with the observed data. The invertedPATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTreflectivity shown in FIG. 4A, together with the history of gradients and models output at each iteration in its derivation, were used to estimate a sparse model representation of the reflectivity with the quadratic constraint to ensure that the data domain match is not compromised, using a small number of iterations. FIG. 4C shows the estimated reflectivity from constrained basis pursuit and least-squares estimated reflectivity, and FIG. 4D shows the observed data overlain with the modelled data using the constrained basis pursuit sparse reflectivity. It can be observed in the constrained basis pursuit sparse reflectivity output that the main peaks are larger, and the oscillating side-lobes are suppressed. This brings the new reflectivity solution closer to the true reflectivity.

[0059] FIG. 5A shows true reflectivity, FIG. 5B shows the LS-RTM recovered model, and FIG. 5C shows the sparse reflectivity derived using the constrained basis-pursuit algorithm with the LI -norm with quadratic constraint. The foregoing results in a sparser reflectivity model without compromising the data fit to within a specified tolerance.

[0060] FIGS. 6A, 6B and 6C show a comparison of true and inverted velocity models as images of spatial distribution of velocity in the Earth’s subsurface. In FIG. 6A, shows the synthetic true-earth P-wave velocity model. FIG. 6B shows the velocity model inverted with FWI using the least-squares based objective function, and FIG. 6C shows the basis pursuit result taking the FWI updated velocity as input and performing basis-pursuit with a total variation objective function with the quadratic constraint thereby promoting sparsity (“blockyness”) in the model. The basis-pursuit result reduces spurious oscillations in the velocity model without compromising the data fit.

[0061] FIG. 7 shows a ID profile through these results. In this case, the TV-norm was used as the sparsity promoting norm rather than the Ll-norm in the basis-pursuit approach. The quadratic constraint was again used to ensure the data fit was not compromised.

[0062] The foregoing process explained with reference to FIG. 2 may be implemented on any form of computer or computer system programmed to perform the actions described. FIG. 8 shows an example computing system 100 in accordance with some embodiments. The computing system 100 may be an individual computer system 101 A or an arrangement of distributed computer systems. The individual computer system 101A may include onePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTor more analysis modules 102 that may be configured to perform various tasks according to some embodiments, such as the tasks explained with reference to FIG. 8. To perform these various tasks, the analysis module 102 may operate independently or in coordination with one or more processors 104, which may be connected to one or more storage media 106. A display device 105 such as a graphic user interface of any known type may be in signal communication with the processor 104 to enable user entry of commands and / or data and to display results of execution of a set of instructions according to the present disclosure.

[0063] The processor(s) 104 may also be connected to a network interface 108 to allow the individual computer system 101A to communicate over a data network 110 with one or more additional individual computer systems and / or computing systems, such as 101B, 101C, and / or 101D (note that computer systems 101B, 101C and / or 101D may or may not share the same architecture as computer system 101 A, and may be located in different physical locations, for example, computer systems 101 A and 10 IB may be at a well drilling location, while in communication with one or more computer systems such as 101C and / or 101D that may be located in one or more data centers on shore, aboard ships, and / or located in varying countries on different continents).

[0064] A processor may include, without limitation, a microprocessor, microcontroller, processor module or subsystem, programmable integrated circuit, programmable gate array, or another control or computing device.

[0065] The storage media 106 may be implemented as one or more computer-readable or machine-readable storage media. Note that while in the example embodiment of FIG. 8 the storage media 106 are shown as being disposed within the individual computer system 101A, in some embodiments, the storage media 106 may be distributed within and / or across multiple internal and / or external enclosures of the individual computing system 101A and / or additional computing systems, e.g., 101B, 101C, 101D. Storage media 106 may include, without limitation, one or more different forms of memory including semiconductor memory devices such as dynamic or static random access memories (DRAMs or SRAMs), erasable and programmable read-only memories (EPROMs),PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTelectrically erasable and programmable read-only memories (EEPROMs) and flash memories; magnetic disks such as fixed, floppy and removable disks; other magnetic media including tape; optical media such as compact disks (CDs) or digital video disks (DVDs); or other types of storage devices. Note that computer instructions to cause any individual computer system or a computing system to perform the tasks described above may be provided on one computer-readable or machine-readable storage medium, or may be provided on multiple computer-readable or machine-readable storage media distributed in a multiple component computing system having one or more nodes. Such computer- readable or machine-readable storage medium or media may be considered to be part of an article (or article of manufacture). An article or article of manufacture can refer to any manufactured single component or multiple components. The storage medium or media can be located either in the machine running the machine-readable instructions, or located at a remote site from which machine-readable instructions can be downloaded over a network for execution.

[0066] It should be appreciated that computing system 100 is only one example of a computing system, and that any other embodiment of a computing system may have more or fewer components than shown, may combine additional components not shown in the example embodiment of FIG. 8, and / or the computing system 100 may have a different configuration or arrangement of the components shown

[0067] in FIG. 8. The various components shown in FIG. 8 may be implemented in hardware, software, or a combination of both hardware and software, including one or more signal processing and / or application specific integrated circuits.

[0068] Further, the acts of the processing methods described above may be implemented by running one or more functional modules in information processing apparatus such as general purpose processors or application specific chips, such as ASICs, FPGAs, PLDs, GPUs, coprocessors or other appropriate devices. These modules, combinations of these modules, and / or their combination with general hardware are all included within the scope of the present disclosure.References Cited in this DisclosurePATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTAragao, O. and Sava, P., 2020, Elastic full-waveform inversion with probabilistic petrophysical model constraints, GEOPHYSICS, 85, 2, R101-R111Deal, M. M., Nolet, G., 1996, Nullspace shuttles, Geophysical Journal International, 124, 372-380.Duveneck, E., Kiehn, M., Chandran, A., and Ktihnel, T., 2021, Reflection angle / azimuth- dependent least-squares reverse time migration, GEOPHYSICS, 86, 5, S325-S338.Fichtner, A., and Zunino, A., 2019, Hamiltonian nullspace shuttles, Geophysical Research Letters, 46, 644-651.Keating, S.D. and Innanen, K.A., 2020, Null-space shuttles for targeted uncertainty analysis in full-waveform inversion, GEOPHYSICS, 86, 1, R63-R76.Keating, S.D. and Innanen, K.A., 2024, Targeted null-space shuttling in time-lapse fullwaveform inversion, GEOPHYSICS, 89, 6, R541-R549.Fehmers, G.C. and C.R.W. Hocker, 2003, Fast structural interpretation with structure- oriented filtering, GEOPHYSICS, 68, 4, 1286-1293.Kamath, N., Tsvankin, I., and Naeini, E.Z., 2017, Faces-constrained fwi: Toward application tor reservoir characterization, The Leading Edge, 36, 11, 924-930 Nemeth, T., Wu, C., and Schuster, G.T., 1999, Least-squares migration of incomplete reflection data, GEOPHYSICS, 64, 1, 208, 221.Newton, I., 1711, De analysi per aequationes numero terminorum infinitas, William Jones, London.Nocedal, J. and Wright, S.J., 2006, Numerical optimization, Springer.Raphson, J., 1690, Analysis Aequationum Universalis, Abel Swale, London.Rocha, D. and Sava, P., 2018, Elastic reflection waveform inversion with petrophysical model constraints, SEG Technical Program Expanded Abstracts, SEG, 1203-1207Singh, S., Tsvankin, L, and Naeini, E.Z., 2018, Bayesian framework for elastic full waveform inversion with facies information, The Leading Edge, 37, 12, 924-931.Tarantola, A., 1984, Inversion of seismic reflection data in the acoustic approximation, GEOPHYSICS, 49, 8, 1259-1266.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPTZhang, Z.D., Alkhalifah, T., Naeini, E.Z., and Sun, B.B., 2018, Multiparameter elastic full waveform inversion with facies-based constraints, Geophysical Journal International, 213, 3, 2112-2127

[0069] In light of the principles and example implementations described and illustrated herein, it will be recognized that the example implementations can be modified in arrangement and detail without departing from such principles. The foregoing discussion has focused on specific implementations, but other configurations are also contemplated. In particular, even though expressions such as in “an implementation," or the like are used herein, these phrases are meant to generally reference implementation possibilities, and are not intended to limit the disclosure to particular implementation configurations. As used herein, these terms may reference the same or different implementations that are combinable into other implementations. As a rule, any implementation referenced herein is freely combinable with any one or more of the other implementations referenced herein, and any number of features of different implementations are combinable with one another, unless indicated otherwise. Although only a few examples have been described in detail above, those skilled in the art will readily appreciate that many modifications are possible within the scope of the described examples. Accordingly, all such modifications are intended to be included within the scope of this disclosure as defined in the following claims.

Claims

PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT CLAIMSWhat is claimed is:

1. A method for estimating the spatial distribution of Earth formation seismic properties in Earth’s subsurface from measurements of seismic signals made proximate a volume of the subsurface to be evaluated, the method comprising:(a) generating an initial model of the spatial distribution and entering the initial model and the measured seismic signals into a computer;(b) in the computer, entering the measured seismic signals and the initial model into an inversion process to obtain an updated model of the spatial distribution;(c) in the computer, estimating a Hessian using information from (b);(d) in the computer, determining a transform of the updated model;(e) in the computer, performing a secondary inversion on a result of (d) with a sparsity promoting norm with a constraint being an approximation of an objective function used in (b) to at least second (quadratic) order to obtain a new transformed updated model that will ensure the quality of the data-domain fit is within a specified tolerance;(f) in the computer, inversely transforming the new transformed updated model; and (g) in the computer, generating an image of the spatial distribution of the seismic properties.

2. The method of claim 1 wherein the inversion that creates the updated model from the initial model comprises at least one of full waveform inversion or least squares migration.

3. The method of claim 1 further comprising determining the Hessian approximation using a combination of data obtained at previous iterations steps in the inversion process.

4. The method of claim 1 wherein sparsity is imposed in a transformed domain.

5. The method of claim 4 wherein the transformed domain comprises, at least one of curvelet, Fourier, Radon, or wavelet domains, or derivatives thereof.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT 6. The method of claim 1 wherein the sparsity with quadratic constraint is imposed in overlapping spatial windows.

7. The method of claim 1 wherein the Hessian approximation is regularised and / or constrained.

8. The method of claim 7 wherein the Hessian approximation is constrained to be positive definite.

9. The method of claim 7 wherein the Hessian approximation is determined using stabilised non- stationary convolution operators.

10. The method of claim 1 wherein a basis pursuit method is applied at or after one or more of the iterations of the inversion process.

11. The method of claim 1 wherein a priori information is used to constrain the additional inversion.

12. The method of claim 1 where the updated model is determined using iterative thresholding in a transformed domain.

13. The method of claim 12 wherein the thresholding is based on the amplitude, energy, and / or phase of the transformed models.

14. The method of claim 1 wherein models and / or gradients generated during the inverting are preconditioned.

15. The method of claim 1 wherein sparsity is imposed using at least one of L0, LI, LP, total variation norms, or combinations thereof.

16. The method of claim 11 where the a priori information comprises a subsurface structural dip.

17. The method of claim 11 where the a priori information comprises data obtained from wells and / or boreholes.PATENT APPLICATION ATTY DOCKET NO. DGL-25-01CPT18. The method of claim 17 further comprising additional constraints including rock-physics or petrophysical constraints.

19. A non-transitory computer readable medium having stored thereon a computer program having logic operable to cause a programmable computer to perform the method according to any of claims 1 through 18.