Multi-scale manifold learning for full waveform inversion

Patent metadata

Field	Value
Publication number	US-10353093-B2
Application number	US-201715661217-A
Country	US
Kind code	B2
Filing date	Jul 27, 2017
Priority date	Jul 27, 2017
Publication date	Jul 16, 2019
Grant date	Jul 16, 2019

Abstract

A method for analyzing acoustic/elastic waves to determine subsurface structure of the earth includes receiving a plurality of observations of a seismic acoustic/elastic wave-field from a plurality of sensors; generating a plurality of Bernstein grids of differing resolutions; calculating a data misfit of the plurality of observations with respect to an initial subsurface structure model defined in terms of conic combinations of Bernstein polynomials on a lowest resolution Bernstein grid, and mapping the data misfit from the Bernstein grid onto a Lagrangian grid; updating the subsurface structure model by minimizing the data misfit between the plurality of observations and observations obtained by a simulation; increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model on the increased resolution Bernstein grid; and mapping the recomputed subsurface structure model onto the Lagrangian grid.

First claim

What is claimed is:

1. A method for analyzing acoustic/elastic waves to determine subsurface structure of the earth, comprising the steps of:
receiving a plurality of observations of a seismic acoustic/elastic wave-field from a plurality of sensors;
generating a plurality of Bernstein grids of differing resolutions;
calculating a data misfit of the plurality of observations with respect to an initial subsurface structure model defined in terms of conic combinations of Bernstein polynomials on a lowest resolution Bernstein grid, and mapping the data misfit from the Bernstein grid onto a Lagrangian grid;
updating the subsurface structure model by minimizing the data misfit between the plurality of observations and observations obtained by a simulation;
increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model on the increased resolution Bernstein grid; and
mapping the recomputed subsurface structure model onto the Lagrangian grid.

2. The method of claim 1 , wherein said steps of updating the subsurface structure model, increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model, and mapping the recomputed subsurface structure model onto the Lagrangian grid are repeated while the Bernstein grid resolution is less or equal than the Lagrangian grid resolution.

3. The method of claim 1 , wherein updating the subsurface structure model by minimizing the data misfit between the plurality of observations and observations obtained by a simulation comprises:
calculating a wave field as a function of time over a finite time interval by running a forward simulation;
calculating an adjoint field as a function of time over the finite time interval by running a backward simulation based on the wave field;
calculating a gradient of the misfit functional from the wave field and the adjoint field, and calculating the misfit from said gradient;
estimating an updated subsurface structure model on the Bernstein grid that reduces the misfit functional itself; and
mapping the updated subsurface structure model onto the Lagrangian grid.

4. The method of claim 3 , wherein said steps of calculating a wave field, calculating an adjoint field, calculating a gradient of the misfit functional and the misfit, estimating an updated subsurface structure model, and mapping the updated subsurface structure model onto the Lagrangian grid are repeated while a stop condition is not satisfied,
wherein said stop condition is |J(m k )−J(m k−1 ))|<Tolerance×J(m*), wherein m k is a current estimate of the subsurface structure model, m k−1 is a previous estimate of the subsurface structure model, J( ) is the misfit functional, m* is a subsurface structure model at a previous resolution of the Bernstein grid, and Tolerance changes on a case-by-case basis.

5. The method of claim 1 , further comprising repeating said steps of minimizing a data misfit of the plurality of observations with respect to an initial subsurface structure model, updating the subsurface structure model, increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model on the increased resolution Bernstein grid; and mapping the recomputed subsurface structure model onto the Lagrangian grid for different initial subsurface structure models, and selecting a subsurface structure model that yields a smallest misfit functional as a best subsurface structure model.

6. The method of claim 3 , wherein:
the initial subsurface structure model is defined as m=Σ i a i B i (x) wherein B i is an i-th Bernstein polynomial defined over the standard element x∈[0,1] 3 and extended to the entire Bernstein grid by continuity, and a i are constants wherein a i ≥0;
wherein calculating a wave field as a function of time over a finite time interval by running a forward simulation solves M(m)ü+F(m, u)=s for a wherein u is the wave field, m is the subsurface structure model defined for points from all Bernstein polynomials in the Bernstein grid, M is a mass matrix, s is seismic source modeled as a point source, F(m, u) is a force wherein F(m, u)=K(m) u where K is a stiffness matrix;
calculating an adjoint field as a function of time over the finite time interval by running a backward simulation based on the wave field solves ä T M+a T K=∂ u r for a wherein a is the adjoint field, a(T)={dot over (a)}(T)=0, u is the wave field, r(u, m, t):=∥d(t)−V u(m, t)∥ 2 , r(x)>0 for x≠0, is a noise model that quantifies the error between the plurality of observations d(t) and the wave field u(m, t) at locations of the plurality of sensors and V is a sampling matrix;
the misfit functional is J(u,m)=Σ s N sources ∫ 0 T r(u, m, t)dt;
the gradient of the misfit functional is ∂ m J=Σ s N sources ∫ 0 T {dot over (β)} dt where {dot over (β)}=(ä T ∂ m M+a T ∂ m K)u+∂ m r, s.t.β(0)=0, and ∂ m K=(K(B 1 ), . . . , K(B n )) are sensitivity matrices.

7. A non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executed by the computer to perform the method steps for analyzing acoustic/elastic waves to determine subsurface structure of the earth, the method comprising the steps of:
receiving a plurality of observations of a seismic acoustic/elastic wave-field from a plurality of sensors;
generating a plurality of Bernstein grids of differing resolutions;
calculating a data misfit of the plurality of observations with respect to an initial subsurface structure model defined in terms of conic combinations of Bernstein polynomials on a lowest resolution Bernstein grid, and mapping the data misfit from the Bernstein grid onto a Lagrangian grid;
updating the subsurface structure model by minimizing the data misfit between the plurality of observations and observations obtained by a simulation;
increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model on the increased resolution Bernstein grid; and
mapping the recomputed subsurface structure model onto the Lagrangian grid.

8. The computer readable program storage device of claim 7 , wherein said steps of updating the subsurface structure model, increasing resolution of the Bernstein grid and recomputing the updated subsurface structure model, and mapping the recomputed subsurface structure model onto the Lagrangian grid are repeated while the Bernstein grid resolution is less or equal than the Lagrangian grid resolution.

9. The computer readable program storage device of claim 7 , wherein updating the subsurface structure model by minimizing the data misfit between the plurality of observations and observations obtained by a simulation comprises:
calculating a wave field as a function of time over a finite time interval by running a forward simulation;
calculating an adjoint field as a function of time over the finite time interval by running a backward simulation based on the wave field;
calculating a gradient of the misfit functional from the wave field and the adjoint field, and calculating the misfit from said gradient;
estimating an updated subsurface structure model on the Bernstein grid that reduces the misfit functional itself; and
mapping the updated subsurface structure model onto the Lagrangian grid.

10. The computer readable program storage device of claim 9 , wherein said steps of calculating a wave field, calculating an adjoint field, calculating a gradient of the misfit functional and the misfit, estimating an updated subsurface structure model, and mapping the updated subsurface structure model onto the Lagrangian grid are repeated while a stop condition is not satisfied,
wherein said stop condition is |J(m k )−J(m k−1 ))|<Toleran

Assignees

• IBM

Inventors

• Akhriev Albert
• Mckenna Sean
• Moore Stephen
• Zhuk Sergiy

Classifications

• G01V2210/6222

  Velocity; travel time · CPC title

• G01V1/282 Primary

  Application of seismic models, synthetic seismograms · CPC title

• G01V2210/679

  Reverse-time modeling or coalescence modelling, i.e. starting from receivers · CPC title

• G01V1/305 Primary

  Travel times · CPC title