Extended Subspace Method for Cross-Talk Mitigation in Multi-Parameter Inversion

Abstract

An extended subspace method for inverting geophysical data to infer models for two or more subsurface physical properties, using gradients of an objective function as basis vectors for forming model updates. The extended set of basis vectors provides explicit mixing between gradient components corresponding to different medium parameters, for example P-wave velocity and an anisotropy parameter. In a preferred embodiment, off-diagonal elements of the mixing matrix may be scaled to adjust the degree of mixing between gradient components. Coefficients of the basis vector expansion are determined in a way that explicitly accounts for leakage or crosstalk between different physical parameters. The same extended subspace approach may be used to make further improvement to the model updates by incorporating well constraints, where well log data are available.

Description

FIELD OF THE INVENTION

This invention relates to the field of geophysical prospecting and, more particularly, to processing geophysical data. Specifically, the invention is a method for inferring properties of the subsurface based on information contained in geophysical data acquired in field experiments.

BACKGROUND OF THE INVENTION

During seismic, electromagnetic, or a similar survey of a subterranean region, geophysical data are acquired typically by positioning a source at a chosen shot location, and measuring seismic, electromagnetic, or another type of back-scattered energy generated by the source using receivers placed at selected locations. The measured reflections are referred to as a single “shot record”. Many shot records are measured during a survey by moving the source and receivers to different locations and repeating the aforementioned process. The survey can then be used to perform inversion, e.g., Full Waveform/Wavefield Inversion in the case of seismic data, which uses information contained in the shot records to determine physical properties of the subterranean region (e.g., speed of sound in the medium, density distribution, resistivity, etc.). Inversion is an iterative process, each iteration comprising the steps of forward modeling to create simulated (model) data and objective function computation to measure the similarity between simulated and field data. Physical properties of the subsurface are adjusted at each iteration to ensure progressively better agreement between simulated and field data. The invention will be described primarily in the context of Full Waveform Inversion of seismic data, but can be applied to inversion of other types of geophysical data.

Multi-parameter inversion involves simultaneous updating of at least two medium properties. A typical strategy is to formulate an objective (cost) function E(m) measuring the misfit between modeled and field data, where m is a vector of medium properties whose components can be compressional and shear-wave velocities, V_p and V_s, density ρ, Thompsen anisotropy parameters ε and δ (Tsvankin, 2001, p. 18), etc. The gradient of the objective function with respect to individual components of m is indicative of the direction in which medium parameters can be updated so that the objective function is minimized and progressively better fit of modeled and field data is obtained. The basis of this approach is the well-known Taylor series:

$E\left(m+\Delta m\right)=E\left(m\right)+\left({\nabla }_mE\right)\Delta m+\frac{1}{2}\Delta m^T\left({\nabla }_{\mathrm{mm}}E\right)\Delta m+\dots ,$

where Am is the desired update; ∇_mE and ∇_mmE are the gradient and the Hessian of the objective function respectively. The gradient ∇_mE is a vector containing first-order derivatives of the objective function E with respect to each individual component m_i of the model vector m:

${\nabla }_mE=\left[\frac{\partial E}{\partial m_i}\right].$

The Hessian ∇_mmE is a matrix containing second-order derivatives of the objective function E with respect to individual components m_i, m_j:

${\nabla }_{\mathrm{mm}}E=\left[\frac{\partial E}{\partial m_i\partial m_j}\right].$

Clearly, if we neglect quadratic terms (the ones with the Hessian) of this expansion and set Δm=−α∇_mE, with α>0, then the objective function will decrease:

E(m+Δm)=E(m)+(∇_mE)Δm=E(m)−α(∇_mE)²<E(m).

Optimal α can be determined with the help of line search, which typically involves evaluating the objective (cost) function for strategically chosen values of α so as to find the best one.

The drawback of this approach is that the gradient does not usually provide the best possible descent direction. Different components of the gradient could be of vastly different magnitudes (especially, when they correspond to different types of medium properties, e.g., V_p and ε) and may exhibit leakage from one component to another due to interdependence of different medium parameters on one another.

A better descent direction can be obtained if the quadratic terms are taken into account. Various approaches of this type are called Newton's method, Newton-CG, and Gauss-Newton and are based on inverting the Hessian:

Δm =−(∇_mmE)⁻¹∇_mE.

Due to its size (typically 10⁹×10⁹ in 3D), the Hessian has to be inverted iteratively, each iteration involving application of the Hessian to a vector. Depending on the problem, the Hessian-vector products (an equivalent term for application of the Hessian to a vector), can be computed analytically, numerically using finite differences, or using the adjoint state method (Heinkenschloss, 2008). Since only a few (usually 10-20) iterations of this iterative process can be afforded in practice, the resulting approximations to the inverse Hessian are usually not very accurate and may not be able to eliminate the leakage (cross-talk) between various medium parameters or provide the correct scaling between different components of the gradient. Moreover, the inversion algorithm may lead to accumulation of artifacts Am, resulting in a suboptimal solution.

A cheaper way to ensure proper relative scaling of the gradient components is to apply the subspace method (Kennett et al., 1988.) The key idea behind this method is to represent the model perturbation as a sum of basis vectors:

Δm=αs₁+βs₂+ . . .

For example, for two different types of medium parameters (e.g., V_p and ε) a customary choice (Sambridge et al., 1991) is:

$\Delta \tilde{m}=\alpha \left[\begin{array}{c}\Delta m_1\\ 0\end{array}\right]+\beta \left[\begin{array}{c}0\\ \Delta m_2\end{array}\right]$

where one typically sets Δm₁˜(−∇_miE), Δm₂˜(−∇_m2E). Δ{tilde over (m)} denotes the updated (improved) model perturbation, as opposed to the original model perturbation

$E\left(m+\Delta \tilde{m}\right)\approx E\left(m\right)+\alpha \left({\nabla }_{m_1}E\right)\Delta m_1+\beta \left({\nabla }_{m_2}E\right)\Delta m_2+\frac{1}{2}\left[\begin{array}{cc}\mathrm{\alpha \Delta }m_1& \mathrm{\beta \Delta }m_2\end{array}\right]\left[\begin{array}{cc}{\nabla }_{m_1m_1}E& {\nabla }_{m_1m_2}E\\ {\nabla }_{m_2m_1}E& {\nabla }_{m_2m_2}E\end{array}\right]\left[\begin{array}{c}\mathrm{\alpha \Delta }m_1\\ \mathrm{\beta \Delta }m_2\end{array}\right]$

Thus, each component of the gradient can be scaled independently so that the resulting search direction is improved. The scaling factors α and β are chosen so that the quadratic approximation to the objective function is minimized:

$\Delta m=\left[\begin{array}{c}\Delta m_1\\ \Delta m_2\end{array}\right].$

It is easy to show that the minimum of the objective function will be obtained if we set

$\left[\begin{array}{c}\alpha \\ \beta \end{array}\right]=-{\left[\begin{array}{cc}\Delta m_1^T\left({\nabla }_{m_1m_1}E\right)\Delta m_1& \Delta m_1^T\left({\nabla }_{m_1m_2}E\right)\Delta m_2\\ \Delta m_2^T\left({\nabla }_{m_2m_1}E\right)\Delta m_1& \Delta m_2^T\left({\nabla }_{m_2m_2}E\right)\Delta m_2\end{array}\right]}^{-1}\left[\begin{array}{c}\left({\nabla }_{m_1}E\right)\Delta m_1\\ \left({\nabla }_{m_2}E\right)\Delta m_2\end{array}\right].$

The cost of determining the values of α and β (which provide the desired scaling of the gradient components) is equal to two applications of the Hessian to a vector (Δm₁ and Δm₂), making this method far cheaper than Newton/Newton-CG/Gauss-Newton.

However, the limitation is that the leakage (cross-talk) cannot be handled effectively, since all the subspace method does is scale each component of the gradient up or down (by α and β).

SUMMARY OF THE INVENTION

In one embodiment, the invention is, referring to the reference numbers in the FIG. 10 flow chart, a computer-implemented method for iteratively inverting measured geophysical data to infer 3D subsurface models of N≧2 physical properties, comprising: (a) providing an initial model (101) for each physical property, wherein a subsurface region is subdivided into discrete cells, each cell having a value of the physical property; (b) for each physical property and for each of a plurality of the cells, representing a search direction (102), indicating whether the initial model needs to be updated positively or negatively, as a linear combination of M>N basis vectors, wherein (bi) each basis vector has its own coefficient in the linear combination, said coefficient to be determined; (bii) each basis vector has a component that is, or is proportional to, a gradient, with respect to model parameters of one of the N physical properties, of an objective function measuring misfit between model-simulated geophysical data and the measured geophysical data; and (biii) the coefficients are simultaneously optimized (103), using a computer, to minimize or maximize the objective function; and (c) using the optimized coefficients to generate search directions (105), and using the search directions to generate an updated model (106) for each physical property.

In a preferred variation of the foregoing embodiment, the degree of mixing between gradient (search direction) components may be adjusted by scaling the off-diagonal components of the mixing matrix, i.e., a matrix whose elements are the coefficients of the basis vector expansion of the search direction.

The dimensionality of the extended subspace of the present invention, i.e., the number of basis vectors M, can in principle be any number greater than N, the number of unknown parameters that are being inverted for. Selecting M=N² allows for leakage between each parameter and all of the others during the inversion process. However, it may be that not all parameters leak into all other parameters. It may be possible to decide based on empirical or theoretical evidence which parameters may potentially have cross-talk among them, and then choose M accordingly. For example, if one is inverting for compressional velocity Vp, shear wave velocity V_s, and anisotropy parameter ε, one might reasonably expect leakage/cross-talk between V_p and V_s, V_p and ε, but not between V_s and ε. So one could have 3 basis vectors for the Vp search direction (gradients w.r.t. V_p, V_s, ε), but only two basis vectors for the V_s and ε search directions, for a total of 7 basis vectors (instead of 9). As an alternative example, one might follow Kennett's approach described above, in which case there would be N(N+1) basis vectors, i.e. 12 for the case of N=3.

In another embodiment of the invention, referring to the flow chart of FIG. 12, the invention is a computer-implemented method for iteratively inverting measured geophysical data to infer 3D subsurface models of N≧2 physical properties, comprising: (a) providing an initial model (121) for each physical property, wherein a subsurface region is subdivided into discrete cells, each cell having a value of the physical property; (b) for each physical property and for each of a plurality of the cells, representing a search direction (122), indicating whether the initial model needs to be updated positively or negatively, as a linear combination of a plurality of basis vectors; (c) determining coefficients (123) of each linear combination by minimizing, using a computer, a difference between one or more true parameters computed from well data or other known subsurface information and corresponding parameters predicted by the updated search direction; and (d) using the determined coefficients to generate search directions (124), and using the search directions to generate an updated model (125) for each physical property.

BRIEF DESCRIPTION OF THE DRAWINGS

The advantages of the present invention are better understood by referring to the following detailed description and the attached drawings, in which:

FIG. 1 shows the true model for the two parameters, V_p (left) and ε (right), used for the test example;

FIG. 2 shows a perfect update Δm (also the “true” search direction) for a given, slightly perturbed (from the true FIG. 1) model;

FIG. 3 simulates a search direction as it might be in actual practice, contaminated with cross-talk between the two parameters;

FIG. 4 shows the same gradient calculations, i.e. search directions, as in FIG. 3, but improved to reduce cross-talk by application of the method of FIG. 10;

FIG. 5 shows initial models to be used in a 2-parameter inversion of synthetic data generated using the true model of FIG. 1;

FIG. 6 shows preconditioned gradients (search directions) of an objective function computing the misfit between the true modeled synthetic data and data simulated using the initial model of FIG. 5;

FIG. 7 shows search directions comparable to those of FIG. 6, but with the gradient computed using the extended subspace method of FIG. 10;

FIG. 8 shows the inverted property models corresponding to FIG. 6;

FIG. 9 shows the inverted property models corresponding to FIG. 7;

FIG. 10 is a flow chart showing basic steps in the present inventive method for using an extended set of basis vectors, with explicit mixing of coefficients, to compute gradient updates to physical property models during iterative inversion of geophysical data;

FIG. 11 is a flow chart showing basic steps for combining the method of FIG. 10 with well (or other known) constraints; and

FIG. 12 is a flow chart showing basic steps for the inversion method of FIG. 11, using well constraints alone, without also using the extended subspace technique of FIG. 10.

The invention will be described in connection with example embodiments.

However, to the extent that the following detailed description is specific to a particular embodiment or a particular use of the invention, this is intended to be illustrative only, and is not to be construed as limiting the scope of the invention. On the contrary, it is intended to cover all alternatives, modifications and equivalents that may be included within the scope of the invention, as defined by the appended claims.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

Extended Subspace Method

The present invention extends the traditional subspace method in a way that explicitly accounts for possible leakage between gradient components. This can be achieved by picking additional basis vectors. Once again, the concept may be illustrated for the case of two different parameters:

$\Delta \tilde{m}={\alpha }_1\left[\begin{array}{c}\Delta m_1\\ 0\end{array}\right]+{\alpha }_2\left[\begin{array}{c}\Delta m_2\\ 0\end{array}\right]+{\beta }_1\left[\begin{array}{c}0\\ \Delta m_1\end{array}\right]+{\beta }_2\left[\begin{array}{c}0\\ \Delta m_2\end{array}\right]$ $E\left(m+\Delta \tilde{m}\right)=E\left(m\right)+\left({\nabla }_{m_1}E\right)\left({\alpha }_1\Delta m_1+{\alpha }_2\Delta m_2\right)+\left({\nabla }_{m_2}E\right)\left({\beta }_1\Delta m_1+{\beta }_2\Delta m_2\right)+\frac{1}{2}\left[\begin{array}{cc}{\alpha }_1\Delta m_1+{\alpha }_2\Delta m_2& {\beta }_1\Delta m_1+{\beta }_2\Delta m_2\end{array}\right]\left[\begin{array}{cc}{\nabla }_{m_1m_1}E& {\nabla }_{m_1m_2}E\\ {\nabla }_{m_2m_1}E& {\nabla }_{m_2m_2}E\end{array}\right]\hspace{1em}\left[\begin{array}{c}{\alpha }_1\Delta m_1+{\alpha }_2\Delta m_2\\ {\beta }_1\Delta m_1+{\beta }_2\Delta m_2\end{array}\right]$

Similarly to the original subspace method, one can get optimal scaling coefficients from:

$\begin{array}{cc}\left[\begin{array}{c}{\alpha }_1\\ {\alpha }_2\\ {\beta }_1\\ {\beta }_2\end{array}\right]=-{\left[\begin{array}{c}\begin{array}{cccc}\Delta m_1^T\left({\nabla }_{m_1m_1}E\right)\Delta m_1& \Delta m_1^T\left({\nabla }_{m_1m_1}E\right)\Delta m_2& \Delta m_1^T\left({\nabla }_{m_1m_2}E\right)\Delta m_1& \Delta m_1^T\left({\nabla }_{m_1m_2}E\right)\Delta m_2\end{array}\\ \begin{array}{cccc}\Delta m_2^T\left({\nabla }_{m_1m_1}E\right)\Delta m_1& \Delta m_2^T\left({\nabla }_{m_1m_1}E\right)\Delta m_2& \Delta m_2^T\left({\nabla }_{m_1m_2}E\right)\Delta m_1& \Delta m_2^T\left({\nabla }_{m_1m_2}E\right)\Delta m_2\end{array}\\ \begin{array}{cccc}\Delta m_1^T\left({\nabla }_{m_1m_2}E\right)\Delta m_1& \Delta m_1^T\left({\nabla }_{m_1m_2}E\right)\Delta m_2& \Delta m_1^T\left({\nabla }_{m_2m_2}E\right)\Delta m_1& \Delta m_1^T\left({\nabla }_{m_2m_2}E\right)\Delta m_2\end{array}\\ \begin{array}{cccc}\Delta m_2^T\left({\nabla }_{m_2m_1}E\right)\Delta m_1& \Delta m_2^T\left({\nabla }_{m_1m_2}E\right)\Delta m_2& \Delta m_2^T\left({\nabla }_{m_2m_2}E\right)\Delta m_1& \Delta m_2^T\left({\nabla }_{m_2m_2}E\right)\Delta m_2\end{array}\end{array}\right]}^{-1}\hspace{1em}\left[\begin{array}{c}{\nabla }_{m_1}E\Delta m_1\\ {\nabla }_{m_1}E\Delta m_2\\ {\nabla }_{m_2}E\Delta m_1\\ {\nabla }_{m_2}E\Delta m_2\end{array}\right]& \left(1\right)\end{array}$

where the superscript T denotes matrix transpose.

The key novelty is that explicit mixing is performed between gradient components corresponding to different medium parameters, e.g., V_p and ε or V_p and ρ. The scaling/mixing coefficients α_i and β_i are determined automatically from Equation 1 at the cost (measured in the number of Hessian applications to a vector) that is equal to the square of the cost of the traditional subspace method. The coefficients α₁ and β₂ are the ones that would have been computed in the traditional subspace method, while α₂ and β₁ correspond to the extended set of basis vectors being introduced in this invention. An important limitation of the method is that curvature information obtained from the Hessian may not be accurate far away from the global minimum, yielding scaling coefficients that would not lead to an improved search direction Δm. Thus, the method as presented so far would be unlikely to work consistently in practice.

Kennett et al. (1988) proposed an alternative approach to selecting an extended set of basis vectors in the subspace method:

$\begin{array}{cc}\Delta \tilde{m}={\alpha }_1\left[\begin{array}{c}\Delta m_1\\ 0\end{array}\right]+{\alpha }_2\left[\begin{array}{c}\left({\nabla }_{m_1m_1}E\right)\Delta m_1\\ 0\end{array}\right]+{\alpha }_3\left[\begin{array}{c}\left({\nabla }_{m_1m_2}E\right)\Delta m_2\\ 0\end{array}\right]+{\beta }_1\left[\begin{array}{c}0\\ \Delta m_2\end{array}\right]+{\beta }_2\left[\begin{array}{c}0\\ \left({\nabla }_{m_2m_1}E\right)\Delta m_1\end{array}\right]+{\beta }_3\left[\begin{array}{c}0\\ \left({\nabla }_{m_2m_2}E\right)\Delta m_2\end{array}\right]& \left(2\right)\end{array}$

However, the cost of this method is much higher (grows as the third power of the cost of the conventional subspace method) due to the need to compute four additional Hessian-vector products. In this case, the matrix in Eqn. (1) would look different, because of the choice of the extended subspace basis vectors. Instead of elements that look like Δm_i^T(∇_mmiE)Δm_j, Eqn. (1) would have Δm_i^T(∇_mimkE)(∇_mkmjE)Δm_j. Therefore not only additional Hessian-vector products would need to be computed, but the matrix of Eqn. (1) would become bigger because there will more basis vectors. In contrast, the present invention's extension of the subspace method utilizes gradients (or vectors obtained from gradients through application of simple processing steps, such as muting, scaling, etc.) with respect to inversion parameters as basis vectors, thus avoiding the need to perform additional Hessian-vector products.

Practical Issues and Further Extensions

The theory underlying the subspace method assumes that the Hessian correctly captures the behavior of the objective function. As mentioned above, when we are dealing with models that are far from the “true” ones, the objective function may not be locally quadratic. In this case Equation (1) may produce inaccurate estimates of α_i and β_i. Moreover, it is customary to replace the Hessian with its “reduced” version—so-called Gauss-Newton Hessian—which itself becomes inaccurate away from the global minimum. Thus, to make the method work in practice, several modifications are helpful.

The first modification is an application of the well-known “trust region” concept. If the values of α_i and β_i turn out to be too large (e.g., requiring a more than 10% update of medium parameters at any given iteration), they need to be scaled down (clipped.) Rewriting the vector of α_i and β_i as a mixing matrix,

$\left[\begin{array}{cc}{\alpha }_1& {\alpha }_2\\ {\beta }_1& {\beta }_2\end{array}\right],$

we can conveniently scale down either row of the matrix, depending on which parameter update exceeds a predefined threshold.

The second modification represents a second key novel step and has to do with adjusting the degree of mixing between gradient (search direction) components. The mixing can be adjusted by scaling the off-diagonal components of the mixing matrix by (γ_α,γ_β):

$\left[\begin{array}{cc}{\alpha }_1& {\gamma }_{\alpha }{\alpha }_2\\ {\gamma }_{\beta }{\beta }_1& {\beta }_2\end{array}\right].$

Then a line search is performed, i.e., evaluate a series of objective functions

$E\left(m+\Delta \tilde{m}\right)=E\left(m+\left[\begin{array}{cc}{\alpha }_1& {\gamma }_{\alpha }{\alpha }_2\\ {\gamma }_{\beta }{\beta }_1& {\beta }_2\end{array}\right]\left[\begin{array}{c}\Delta m_1\\ \Delta m_2\end{array}\right]\right)$

and select the values of (γ_α,γ_β) corresponding to the best (i.e. minimum or maximum, depending upon how the objective function is formulated) objective function. (Note that the γ_i are introduced for convenience; we could just as well have found optimal values of the off-diagonal elements of the mixing matrix). There are many known ways to perform the line search, but for purposes of the present invention, in order to minimize the computation cost, it is preferable to fit a quadratic form in (γ_α,γ_β) to the objective function above and then find optimal values of (γ_α,γ_β):

E(m+Δ{tilde over (m)};γ_αγ_β)=α₀+α₁γ_α+α₂γ_β+α₃γ_α²+α₄γ_β²+α₅γ_αγ_β.

The objective function is evaluated at six different points (γ_αγ_β), e.g., (1,1), (0.75,1), (1,0.75), (0.5,1), (1,0.5), (0.5,0.5) and the resulting system of linear equations solved for α_i. When the quadratic form is not positive definite, and end point (either 0 or 1) can be chosen for each γ. Note that this line search is different from the traditional one and serves a different purpose. Conventionally, the line search is performed to determine the best possible step size (scaling of the model update), while it is used here to determine the best possible set of mixing coefficients that minimize leakage/cross-talk between different inversion parameters. Once the mixing coefficients are determined and updated search directions are obtained, a conventional line search can be applied to further scale the updated search directions.

The third key novel step addresses the situation in which the level of cross-talk is spatially varying, so that scaling factors (γ_α,γ_β) need to be spatially varying as well. The line search can be performed separately for each shot, producing a spatially varying set of scaling factors. Note that the cost of performing the line search for each shot individually is the same as the cost of traditional spatially invariant line search. The only difference is that instead of summing all individual objective functions computed for each shot record and then selecting the values of (γ_α,γ_β) that correspond to the best cumulative objective function, the selection is performed shot-by-shot, skipping the summation. Each shot is assigned a spatial location and the selected optimal value of (γ_α,γ_β) is also assumed to occur at that location. Finally, interpolation may be performed to obtain a spatially varying distribution of optimal scaling factors (γ_α,γ_β), followed by optional smoothing to avoid introducing artifacts into the inversion. FIG. 10 is a self-explanatory flowchart showing basic steps in this embodiment of the present inventive method.

Incorporating Well Constraints

The idea of using gradients as basis vectors for forming an improved update (search direction) in inversion can be extended to the case in which well logs or other reliable information regarding the subsurface is available, representing another key novel step. Similarly to the methodology described in the previous sections, an improved update (search direction) can be obtained by setting

Δ{tilde over (m)}_i=wⁱ₁Δm₁+wⁱ₂Δm₂+wⁱ₃e  (3),

where i=1,2; e is a vector with all components set to “1”. The unknown coefficients wⁱ₁, wⁱ₂, wⁱ₃ can be determined by requiring that the improved model update fit the “true” well-log-based update

Δm_i^true=m_i^welllog−m_i^current

in some norm:

∥Δm_i^true−Δm_i∥_Ln→min.

In general, optimal coefficients wⁱ₁, wⁱ₂, wⁱ₃ can be found numerically. If n=2, i.e., the L₂ norm is used, the solution to this minimization problem is given by

$\begin{array}{cc}\left[\begin{array}{c}{w}_1^i\\ w_2^i\\ w_3^i\end{array}\right]={\left[\begin{array}{ccc}\Delta m_1^T\Delta m_1& \Delta m_1^T\Delta m_2& \Delta m_1^Te\\ \Delta m_2^T\Delta m_1& \Delta m_2^T\Delta m_2& \Delta m_2^Te\\ e^T\Delta m_1& e^T\Delta m_2& e^Te\end{array}\right]}^{-1}\left[\begin{array}{c}\Delta m_1^T\Delta m_i^{\mathrm{true}}\\ \Delta m_1^T\Delta m_i^{\mathrm{true}}\\ e^T\Delta m_i^{\mathrm{true}}\end{array}\right]& \left(4\right)\end{array}$

The Δm_i can be set proportional (or equal) to the gradients of E, their preconditioned/modified versions, or the improved search directions coming from the extended subspace method described in the previous sections.

There are two key differences with the extended subspace method described previously. First, there is effectively no additional computational cost to be incurred in computing an improved search direction based on the well log information because Hessian-vector products need not be computed and just a small 3×3 matrix has to be inverted. Secondly, the set of basis vectors was extended even further by including vector e. This vector allows us to determine the background (“DC”) component of the update. It is well known that FWI cannot correctly compute the background update when seismic data are missing low frequencies, as is the case for most datasets acquired to date. For some parameters, such as Thompsen's anisotropy parameter δ, this is impossible under any circumstances based on surface seismic data alone. Thus, the vector e was not included previously because it would have been difficult to obtain it reliably. (The availability of a direct measurement of subsurface medium parameters at well locations changes the situation.) Of course, e can be more general than a vector consisting of “1”. For example, it could be a depth-varying function.

If more than one well is available, optimal coefficients wⁱ₁, wⁱ₂, wⁱ₃ should preferably be found at each well location and spatially interpolated between wells and extrapolated away from the wells.

In a typical application the extended subspace method based on the surface seismic data might be used first to produce an improved model update, i.e. search direction, followed by a further modification based on the well log information. Basic steps in this embodiment of the invention are shown in the self-explanatory FIG. 11. Note that not all parameters may be constrained by either well logs or surface seismic data, so the two steps (extended subspace and well constraints) need not apply to the same set of parameters. For example, one could compute improved search directions for V_p and ε based on the Hessian of the objective functions and then compute improved search directions for V_p (again) and δ based on the well log information.

Additionally, application of the extended subspace method could be skipped and well log information used directly to obtain an improved search direction. Basic steps in this embodiment of the invention are shown in the self-explanatory FIG. 12. The advantage of this approach is that the significant computation cost associated with the evaluation of the Hessian-vector products required by Equation 1 is avoided. Furthermore, since well information represents a measurement of the actual subsurface properties, the updated search directions can be considered to be optimal and the traditional line search performed to determine optimal scaling of model updates can be skipped as well. The implication of this choice is that the model update no longer relies directly on the assumption that the objective (cost) function value should improve at each iteration. It is entirely possible that the fit between simulated and field data may temporarily become worse, although model fit (i.e., how closely the model approximates subsurface properties) gets better. This situation is known as “local minimum”, reflecting the fact that the objective function may go through peaks and troughs as we progress from the initial model to the true one, and reaches its overall optimal value (“global minimum”) only at the end of the process. Conventional derivative-based methods are not able to overcome the “local minimum” problem, so incorporating well log information and skipping traditional line search may allow the inversion to converge to a significantly better model.

EXAMPLES

The present inventive method was tested using synthetic data generated by assuming the “true” models for the parameters V_p and ε shown in FIG. 1, based on the SEAM model. (See “SEAM update: Completion of Phase I Acoustic Simulations,” The Leading Edge, June, 2010.) First, a model was chosen (not shown) with a very small perturbation in both for V_p and ε to ensure that the Gauss-Newton Hessian is a good measure of the curvature of the objective function. The slightly perturbed model represents a typical current model in the course of iterative inversion. The “true” search directions were then constructed, which were the difference between the true and the perturbed models (FIG. 2). As an illustration of the cross-talk that could be contaminating search directions in a realistic inversion, a linear combination of the “true” search directions was computed, thereby introducing low-frequency overprint into the search direction for V_p and high-frequency reflectivity overprint into the search direction for ε (FIG. 3). “De-mixing” (solving for the coefficients α₁, α₂, β₁, and β₂) was then performed using the extended subspace method of the present invention (Equation 1), which was able to remove the low-frequency overprint from the V_p search direction and almost remove the high-frequency overprint from the ε search direction (FIG. 4).

Next, a two-parameter inversion was performed for V_p and ε using the initial model shown in FIG. 5. Comparing the true and initial models (FIGS. 1 and 5), it can be seen that the update for V_p should be dominated by reflectivity, while the update for ε should be smooth. Due to the cross-talk between V_p and ε, the gradient for V_p contains an undesirable low-frequency component (FIG. 6, oval at left), which may prevent inversion from converging to the correct solution. Computing the de-mixing coefficients using Eqn. 1 and using them to compute new search directions for V_p and ε, shown in FIG. 7, we are able to reduce the undesirable low frequency content in the V_p search direction, while enhancing the corresponding component of the ε search direction (see ovals). As mentioned above, this transformation is performed at each iteration of the inversion. (The gradient shown in FIGS. 6 and 7 were preconditioned, which means that the raw gradients were gained in depth to compensate for the decay of wavefields as they propagate down.) Finally, FIGS. 8 and 9 illustrate that the application of the extended subspace method leads to a better inversion result. The oval in the FIG. 8 ε model indicates regions, where ε has not been recovered correctly, leading to mispositioning of the reflectors (indicated by the arrow) in the V_p model. The oval in the FIG. 9 ε model indicates regions, where E has been recovered better than before, leading to correct positioning of the reflectors (indicated by the arrow) in the V_p model. FIG. 8 was generated using the Kennett subspace method. The inversion results of both figures are after several iteration cycles.

The foregoing description is directed to particular embodiments of the present invention for the purpose of illustrating it. It will be apparent, however, to one skilled in the art, that many modifications and variations to the embodiments described herein are possible. All such modifications and variations are intended to be within the scope of the present invention, as defined by the appended claims.

REFERENCES

• 1. Heinkenschloss, M., “Numerical Solution of Implicitly Constrained Optimization Problems,” CAAM Technical Report TR08-05 (2008). http://www.caam.rice.edu/˜heinken/papers/MHeinkenschloss_—2008a.pdf
• 2. Kennett, B. L. N., Sambridge, M. S., Williamson, P. R., “Subspace methods for large scale inverse problems involving multiple parameter classes,” Geophysical Journal International 94, 237-247 (1988).
• 3. Sambridge, M. S., Tarantola, A., Kennett, B. L. N., “An alternative strategy for non-linear inversion of seismic waveforms,” Geophysical Prospecting 39, 723-736 (1991).
• 4. Tsvankin, I., Seismic Signatures and Analysis of Reflection Data in Anisotropic Media, Pergamon Press, page 18 (2001).

Claims

1. A computer-implemented method for iteratively inverting measured geophysical data to infer 3D subsurface models of N≧2 physical properties, comprising: providing an initial model for each physical property, wherein a subsurface region is subdivided into discrete cells, each cell having a value of the physical property;for each physical property and for each of a plurality of the cells, representing a search direction, indicating whether the initial model needs to be updated positively or negatively, as a linear combination of M>N basis vectors, wherein: each basis vector has its own coefficient in the linear combination, said coefficient to be determined;the basis vectors are chosen such that their coefficients account for cross-talk between the N physical properties during the inversion; andoptimal values of the coefficients are simultaneously solved for, using a computer, to minimize or maximize an objective function measuring misfit between model-simulated geophysical data and the measured geophysical data, wherein solving for the coefficients is performed at least by numerical computation constrained by the geophysical data; andusing the optimal coefficients to generate search directions, and using the search directions to generate an updated model for each physical property.

2. The method of claim 1, wherein each basis vector has a component that is, or is proportional to, a gradient, with respect to model parameters of one of the N physical properties, of the objective function.

3. The method of claim 1, wherein a mixing matrix is formed from the coefficients of the basis vectors, and degree of mixing between the model updates for the N physical properties is adjusted by scaling off-diagonal elements of the mixing matrix with scaling factors that are optimized in the inversion.

4. The method of claim 3, wherein the optimizing of the scaling factors is based on a line search.

5. The method of claim 4, wherein the line search is performed by steps comprising fitting a polynomial function of the scaling factors to the objective function and then finding values of the scaling factors that optimize the objective function.

6. The method of claim 4, further comprising performing a conventional line search to determine an optimal step size for the model update.

7. The method of claim 3, wherein the scaling factors are spatially dependent.

8. The method of claim 7, wherein the optimization comprises performing line searches, and a separate line search is performed for each source shot in a geophysical survey that generated the measured geophysical data, thereby providing the spatial dependence.

9. The method of claim 3, wherein the optimization of the coefficients and the scaling factors is based on one or more Hessians of the objective function.

10. The method of claim 3, wherein optimal values of the scaling factors are determined by a line search comprising evaluating the objective function E for an updated model m+Δm for a plurality of different values of the scaling factors, and selecting a combination of scaling factors giving a least value of E.

11. The method of claim 1, further comprising adjusting the search directions before the generating an updated model for each physical property, wherein each adjusted search direction is represented by a linear combination of a plurality of basis vectors, and coefficients of the linear combination are determined by minimizing a difference between one or more true parameters computed from well data or other known subsurface information and corresponding parameters predicted by the updated search direction.

12. The method of claim 11, wherein the plurality of basis vectors is N+1 in number, comprising a gradient of the objective function with respect to model parameters of each of the N physical properties, plus an additional basis vector whose coefficient allows matching to the well data or other known subsurface information.

13. The method of claim 12, wherein every component of the additional basis vector is unity.

14. The method of claim 1, wherein the geophysical data are seismic data, and the N physical properties are selected from the group consisting of compressional and shear-wave velocities, Vp and Vs, density ρ, and Thompsen anisotropy parameters ε and δ.

15. The method of claim 1, wherein the optimization is based on one or more Hessians of the objective function.

16. A computer-implemented method for iteratively inverting measured geophysical data to infer 3D subsurface models of N≧2 physical properties, comprising: providing an initial model for each physical property, wherein a subsurface region is subdivided into discrete cells, each cell having a value of the physical property;for each physical property and for each of a plurality of the cells, representing a search direction, indicating whether the initial model needs to be updated positively or negatively, as a linear combination of a plurality of basis vectors;determining coefficients of each linear combination by minimizing, using a computer, a difference between one or more true parameters computed from well data or other known subsurface information and corresponding parameters predicted by the updated search direction; andusing the determined coefficients to generate search directions, and using the search directions to generate an updated model for each physical property.

17. The method of claim 16, wherein the plurality of basis vectors is N+1 in number, comprising a gradient of the objective function with respect to model parameters of each of the N physical properties, plus an additional basis vector whose coefficient allows matching to the well data or other known subsurface information.

18. The method of claim 17, wherein every component of the additional basis vector is unity.

19. The method of claim 1, wherein rock-physics-based or empirical relationships between physical property parameters, or well data constraints, or both, are used to reduce number of search direction coefficients to be solved for by the iterative numerical computation constrained by the geophysical data.