Reduced Order Model Based Non-Linear Preconditioning For Iterative Seismic Data Interpretation Via Full Waveform Inversion
Publication Date: 2015-Jun-11
The IP.com Prior Art Database
Page 01 of 6
REDUCED ORDER MODEL BASED NON-LINEAR PRECONDITIONING FOR ITERATIVE SEISMIC DATA INTERPRETATION VIA FULL WAVEFORM INVERSION
Seismic data interpretation based on iterative fitting of the measured data with a PDE wave propagation model known as the full waveform inversion (FWI) is the most general yet a very computationally challenging approach. The difficulties arise from the fact that when formulated as non-linear least squares, it yields a highly non-linear optimization problem with a functional that has many local minima. Thus, any local derivative-based method has difficulties obtaining a global minimizer. Even when convergence is achieved, it typically needs many iterations. Since each iteration requires one or more solutions of a large scale wave propagation problem, the overall cost of FWI can become prohibitively expensive. In our novel approach we replace the traditional least squares functional with a non-linearly preconditioned one. The preconditioner maps the data space to the reduced order model (ROM) space, so instead of minimizing the data misfit the optimization seeks a minimizer of a misfit between the ROMs. Since the ROMs are closely related to the full wave propagation models, the optimization functional becomes much better behaved than the standard one. This leads to a robust convergence from poor starting models and to a reduction in the total number of iterations. Our approach is general enough to be applicable to a variety of time domain wave inversion problems. This includes surface and borehole (VSP and cross-well) seismics, acoustic (sonic) logging, etc.
Page 02 of 6
Time domain seismic data interpretation via FWI is typically formulated as a non- linear least squares optimization problem
minimize ‖𝒅 − 𝓕[𝒎]‖𝟐𝟐 + 𝜶𝓡(𝒎),
where the data measured at receiver locations is 𝑑, the model is 𝑚, the forward modeling operator ℱ maps the model space to the data space, and the regularization functional ℛ is weighted by a penalty parameter 𝛼. The minimization problem is then solved by a derivative-based method such as the steepest descent, non-linear conjugate gradients (NLCG), Gauss-Newton (GN) or quasi-Newton (BFGS, L-BFGS) .
The highly non-linear nature of the minimization problem (1) makes it hard to solve numerically since the derivative-based methods may get stuck in local minima due to cycle skipping, noisy data, etc. A way to alleviate these issues is to modify the optimization functional in (1) by employing a preconditioner. Commonly, the preconditioned functional takes the form
𝓒(𝐦) = 𝚫𝒅∗ 𝑾 𝚫𝒅 + 𝜶𝓡(𝒎),
where Δ𝑑 = 𝑑 − ℱ[𝑚] and 𝑊 is a weighting matrix . Other approaches rely on a particular optimization method. For example, a conjugate gradient method is often preconditioned with multiplying the gradient by an approximation to the diag...