William W. Symes
An extremely simple single-trace transmission example shows how an extended source formulation of full waveform inversion can produce an optimization problem without spurious local minima ("cycle skipping"). The data consist of a single trace recorded at a given distance from a point source. The velocity or slowness is presumed homogeneous, and the target source wavelet is presumed quasi-impulsive or focused at zero time lag. The source is extended by permitting energy to spread in time, and the spread is controlled by adding a weighted mean square of the extended source wavelet to the data misfit, to produce the extended inversion objective. The objective function and its gradient can be computed explicitly, and it is easily seen that all local minimizers must be within a wavelength of the correct slowness. The derivation shows several important features of all similar extended source algorithms. For example, nested optimization, with the source estimation in the inner optimization (variable projection method), is essential. The choice of the weight operator, controlling the extended source degrees of freedom, is critical: the choice presented here is a differential operator, and that property is crucial for production of an objective immune from cycle-skipping.
Kirk D. Blazek, Christiaan C. Stolk, William W. Symes
This paper provides a theoretical foundation for some common formulations of inverse problems in wave propagation, based on hyperbolic systems of linear integro-differential equations with bounded and measurable coefficients. The coefficients of these time-dependent partial differential equations respresent parametrically the spatially varying mechanical properties of materials. Rocks, manufactured materials, and other wave propagation environments often exhibit spatial heterogeneity in mechanical properties at a wide variety of scales, and coefficient functions representing these properties must mimic this heterogeneity. We show how to choose domains (classes of nonsmooth coefficient functions) and data definitions (traces of weak solutions) so that optimization formulations of inverse wave problems satisfy some of the prerequisites for application of Newton's method and its relatives. These results follow from the properties of a class of abstract first-order evolution systems, of which various physical wave systems appear as concrete instances. Finite speed of propagation for linear waves with bounded, measurable mechanical parameter fields is one of the by-products of this theory.
William W. Symes, Huiyi Chen, Susan E. Minkoff
Study of a simple single-trace transmission example shows how an extended source formulation of full-waveform inversion can produce an optimization problem without spurious local minima ("cycle skipping"), hence efficiently solvable via Newton-like local optimization methods. The data consist of a single trace extracted from a causal pressure field, propagating in a homogeneous fluid according to linear acoustics, and recorded at a given distance from a transient point energy source. The source intensity ("wavelet") is presumed quasi-impulsive, with zero energy for time lags greater than a specified maximum lag. The inverse problem is: from the recorded trace, recover both the sound velocity or slowness and source wavelet with specified support, so that the data is fit with prescribed RMS relative error. The least-squares objective function has multiple large residual minimizers. The extended inverse problem permits source energy to spread in time, and replaces the maximum lag constraint by a weighted quadratic penalty. A companion paper shows that for proper choice of weight operator, any stationary point of the extended objective produces a good approximation of the global minimizer of the least squares objective, with slowness error bounded by a multiple of the maximum lag and the assumed noise level. This paper summarizes the theory developed in the companion paper and presents numerical experiments demonstrating the accuracy of the predictions in concrete instances. We also show how to dynamically adjust the penalty scale during iterative optimization to improve the accuracy of the slowness estimate.
William W. Symes
Source extension is a reformulation of inverse problems in wave propagation, that at least in some cases leads to computationally tractable iterative solution methods. The core subproblem in all source extension methods is the solution of a linear inverse problem for a source (right hand side in a system of wave equations) through minimization of data error in the least squares sense with soft imposition of physical constraints on the source via an additive quadratic penalty. A variant of the time reversal method from photoacoustic tomography provides an approximate solution that can be used to precondition Krylov space iteration for rapid convergence to the solution of this subproblem. An acoustic 2D example for sources supported on a surface, with a soft contraint enforcing point support, illustrates the effectiveness of this preconditioner.
William W. Symes
Nonlinear least squares data-fitting driven by physical process simulation is a classic and widely successful technique for the solution of inverse problems in science and engineering. Known as "Full Waveform Inversion" in application to seismology, it can extract detailed maps of earth structure from near-surface seismic observations, but also suffers from a defect not always encountered in other applications: the least squares error function at the heart of this method tends to develop a high degree of nonconvexity, so that local optimization methods (the only numerical methods feasible for field-scale problems) may fail to produce geophysically useful final estimates of earth structure, unless provided with initial estimates of a quality not always available. A number of alternative optimization principles have been advanced that promise some degree of release from the multimodality of Full Waveform Inversion, amongst them Wavefield Reconstruction Inversion, the focus of this paper. Applied to a simple 1D acoustic transmission problem, both Full Waveform and Wavefield Reconstruction Inversion methods reduce to minimization of explicitly computable functions, in an asymptotic sense. The analysis presented here shows explicitly how multiple local minima arise in Full Waveform Inversion, and that Wavefield Reconstruction Inversion can be vulnerable to the same "cycle-skipping" failure mode.
William W. Symes
A simple inverse problem for the wave equation requires determination of both the wave velocity in a homogenous acoustic material and the transient waveform of an isotropic point radiator, given the time history of the wavefield at a remote point in space. The duration (support) of the source waveform and the source-to-receiver distance are assumed known. A least squares formulation of this problem exhibits the "cycle-skipping" behaviour observed in field scale problems of this type, with many local minima differing greatly from the global minimizer. An extended formulation, dropping the support constraint on the source waveform in favor of a weighted quadratic penalty, eliminates this misbehaviour. With proper choice of the weight operator, the velocity component at any local minimizer of this extended objective function differs from the global minimizer of the least-squares formulation by less than a linear combination of the source waveform support radius and data noise-to-signal ratio.
William W. Symes
Adaptive Waveform Inversion (AWI) applied to transient transmitted wave data can yield estimates of index of refraction (or wave velocity) similar to those obtained by travel time inversion. The AWI objective function measures normalized mean-square dispersion about zero time of a family of filters, one filter for each source-reeciver pair, designed to match predicted data to observed data. Provided that the data contain a single smooth wavefront, this function approaches the mean square traveltime error as data wavelength tends to zero. The scaling of each filter to have unit norm is responsible for the simple relation with travel time tomography. A similar approach, Matched Source Waveform Inversion (MSWI), does not normalize the filters and has a looser relation with mean-square travel time error. If substantial energy is spread amongst multiple arrivals, on the other hand, neither AWI nor MSWI objectives approximate the travel time mean square erorr, and attempts to minimize them by local (Newton-related) optimization are as likely to stagnate at models predicting erroneous travel times as is least-square Full Waveform Inversion (FWI). The matching filter at the heart of this approach must be ``pre-whitened'', that is, computed by solving a regularized least squares problem for filtered data misfit. In order for the relation to traveltime tomography to hold, and for the filtered data misfit to stay relatively small, the regularization weight must be coupled to data wavelength. These objective functions can also be used as penalty terms in a homotopy of objectives connecting AWI or MSWI with FWI.
William W. Symes, Huiyi Chen, Susan E. Minkoff
Iterative inversion of seismic, ultrasonic, and other wave data by local gradient-based optimization of mean-square data prediction error (Full Waveform Inversion or FWI) can fail to converge to useful model estimates if started from an initial model predicting wave arrival times in error by more than half a wavelength (a phenomenon known as cycle skipping). Matched Source Waveform Inversion (MSWI) extends the wave propagation model by a filter that shifts predicted waves to fit observed data. The MSWI objective adds a penalty for deviation of this filter from the identity to the mean-square data misfit . The extension allows the inversion to make large model adjustments while maintaining data fit and so reduces the chances of local optimization iterates stagnating at non-informative model estimates. Theory suggests that MSWI applied to acoustic transmission data with single-arrival wavefronts may produce an estimate of refractive index similar to the result of travel time inversion, but without requiring explicit identification of travel times. Numerical experiments conform to this expectation, in that MSWI applied to single arrival transmission data gives reasonable model estimates in cases where FWI fails. This MSWI model can then be used to jumpstart FWI for further refinement of the model. The addition of moderate amounts of noise (30\%) does not negatively impact MSWI's ability to converge. However, MSWI applied to data with multiple arrivals is no longer theoretically equivalent to travel-time tomography and exhibits the same tendency to cycle-skip as does FWI.
Gaoming Chen, Fadil Santosa, William W. Symes
Professor Pierre Sabatier contributed much to the study of inverse problems in theory and practice. Two of these contributions were a focus on theory that actually supports practice, and the identification of well-posed aspects of inverse problems that may quite ill-posed. This paper illustrates these two themes in the context of Electrical Impedance Tomography (EIT), which is both very ill-posed and very practical. We show that for a highly constrained version of this inverse problem, in which a small elliptical inclusion in a homogeneous background is to be identified, optimization of the experimental design (that is, electrode locations) vastly improves the stability of the solution.