Seismic waveforms contain much information that is ignored under standard processing schemes; seismic waveform inversion seeks to use the full information content of the recorded wavefield. In this paper I present, apply, and evaluate a frequency-space domain approach to waveform inversion. The method is a local descent algorithm that proceeds from a starting model to refine the model in order to reduce the waveform misfit between observed and model data. The model data are computed using a full-wave equation, viscoacoustic, frequencydomain, finite-difference method. Ray asymptotics are avoided, and higher-order effects such as diffractions and multiple scattering are accounted for automatically. The theory of frequency-domain waveform/wavefield inversion can be expressed compactly using a matrix formalism that uses finite-difference/finite-element frequency-domain modeling equations. Expressions for fast, local descent inversion using back-propagation techniques then follow naturally. Implementation of these methods depends on efficient frequency-domain forward-modeling solutions; these are provided by recent developments in numerical forward modeling. The inversion approach resembles prestack, reverse-time migration but differs in that the problem is formulated in terms of velocity (not reflectivity), and the method is fully iterative. I illustrate the practical application of the frequencydomain waveform inversion approach using tomographic seismic data from a physical scale model. This allows a full evaluation and verification of the method; results with field data are presented in an accompanying paper. Several critical processes contribute to the success of the method: the estimation of a source signature, the matching of amplitudes between real and synthetic data, the selection of a time window, and the selection of suitable sequence of frequencies in the inversion. An initial model for the inversion of the scale model data is provided using standard traveltime tomographic methods, which provide a robust but low-resolution image. Twenty-five iterations of wavefield inversion are applied, using five discrete frequencies at each iteration, moving from low to high frequencies. The final results exhibit the features of the true model at subwavelength scale and account for many of the details of the observed arrivals in the data.

Prestack migration and/or inversion may be implemented in either the time or the frequency domain. In the frequency domain, it is possible to discretize the frequencies with a much larger sampling interval than that dictated by the sampling theorem and still obtain an imaging result that does not suffer from aliasing (wrap around) in the depth domain. The selection of input frequencies can be reduced when a range of offsets is available; this creates a redundancy of information in the wavenumber coverage of the target. In order to optimize the use of this information, we define a new discretization strategy that depends on the maximum effective offset present in the surface seismic survey: the larger the range of offsets, the fewer frequencies are required.The strategy, exact in a homogeneous 1D earth, selects frequencies by making use of the well-known effect of image stretch in normal-moveout (NMO) correction and in migration (usually considered detrimental for the imaging). The strategy is also useful in more general earth models: we apply it to the 2D Marmousi model and recover a continuous range of wavenumbers using only three input frequencies. The Marmousi inversion result accurately predicts all other data frequencies, demonstrating the redundancy of the data.

Frequency‐domain methods are well suited to the imaging of wide‐aperture cross‐hole data. However, although the combination of the frequency domain with the wavenumber domain has facilitated the development of rapid algorithms, such as diffraction tomography, this has also required linearization with respect to homogeneous reference media. This restriction, and association restrictions on source‐receiver geometries, are overcome by applying inverse techniques that operate in the frequency‐space domain. In order to incorporate the rigorous modelling technique of finite differences into the inverse procedure a nonlinear approach is used. To reduce computational costs the method of finite differences is applied directly to the frequency‐domain wave equation. The use of high speed, high capacity vector computers allow the resultant finite‐difference equations to be factored in‐place. In this way wavefields can be computed for additional source positions at minimal extra cost, allowing inversions to be generated using data from a very large number of source positions. Synthetic studies show that where weak scatter approximations are valid, diffraction tomography performs slightly better than a single iteration of non‐linear inversion. However, if the background velocities increase systematically with depth, diffraction tomography is ineffective whereas non‐linear inversion yields useful images from one frequency component of the data after a single iteration. Further synthetic studies indicate the efficacy of the method in the time‐lapse monitoring of injection fluids in tertiary hydrocarbon recovery projects.

S U M M A R YModern wide-angle surveys are often multi-fold and multi-channel, with densely sampled source and receiver spacings. Such closely spaced data are potentially amenable to multi-channel techniques involving wavefield propagation methods, such as those commonly used in reflection data processing. However, the wide-angle configuration requires techniques capable of handling very general wave types, including those not commonly used in reflection seismology. This is a situation analogous to that faced in cross-borehole seismics, where similar wave types are also recorded. In a real crossborehole example, we compare pre-stack migration, traveltime tomography and wavefield inversion. We find that wavefield inversion produces images that are quantitative in velocity (as are the tomograms) but are of significantly higher resolution; the wavefield inversion results have a resolution comparable to that of the (qualitative) pre-stack migration images. We seek to extend this novel development to the largerscale problem of crustal imaging.An essential element of the approach we adopt is its formulation entirely within the temporal frequency domain. This has three principal advantages: (1) we can choose to 'decimate' the data, by selecting only a limited number of frequency components to invert, thus making inversion of data from large numbers of source positions feasible;(2) we can mitigate the notorious non-linearity of the seismic inverse problem by progressing from low-frequency components in the data to high-frequency components; and (3) we can include in the model any arbitrary frequency dependence of inelastic attenugtion factors, Q(w), and indeed solve for the spatial distribution of Q.An initial synthetic test with an anomaly located within the middle crust yields a velocity image with the correct structural features of the anomaly and the correct magnitude of velocity anomaly. This is related to the fact that the reconstruction is obtained from forward-scattered waves. Under these conditions, the method thus behaves much like tomography. A second test with a deeper, more extensive anomaly yields an image with the correct velocity polarity and the correct location, but with a deficiency in low and high wavenumbers. In this case, this is because the reconstruction is obtained from backscattered waves; under these conditions the method behaves not like tomography, but like migration.A more extensive test, based on a large wide-angle survey in south-eastern California and western Arizona, demonstrates a real potential for high-resolution imaging of crustal structures. Although our results are limited by the acoustic approximation and by the relatively low frequencies that we can model today, the images are sufficiently encouraging to warrant future research. The problem of local minima in the objective function is the most significant practical problem with our method, but we propose that appropriate 'layer' stripping methods can handle this problem.

SUMMARY A comprehensive validation of 2‐D, frequency‐domain, acoustic wave‐equation tomography was undertaken in a ‘blind test’, using third‐party, realistic, elastic wave‐equation data. The synthetic 2‐D, wide‐angle seismic data were provided prior to a recent workshop on the methods of controlled source seismology; the true model was not revealed to the authors until after the presentation of our waveform tomography results. The original model was specified on a detailed grid with variable P‐wave velocity, S‐wave velocity, density and viscoelastic Q‐factor structure, designed to simulate a section of continental crust 250 km long and 40 km deep. Synthetic vertical and horizontal component data were available for 51 shot locations (spaced every 5 km), recorded at 2779 receivers (spaced every 90 m), evenly spread along the surface of the model. The data contained energy from 0.2 to 15 Hz. Waveform tomography, a combination of traveltime tomography and 2‐D waveform inversion of the early arrivals of the seismic waveforms, was used to recover crustal P‐velocity structure from the vertical component data, using data from 51 sources, 1390 receivers and frequencies between 0.8 and 7.0 Hz. The waveform tomography result contained apparent structure at wavelength‐scale resolution that was not evident on the traveltime tomography result. The predicted (acoustic) waveforms in the final result matched the original elastic data to a high degree of accuracy. During the workshop, the exact model was revealed; over much of the model the waveform tomography results provided a good correspondence with the true model, from large‐ to intermediate‐(wavelength) scales, with a resolution limit on the order of 1 km. A significant, near‐surface low‐velocity zone, invisible to traveltime methods, was correctly recovered; the results also provided a high‐resolution image of the complex structure of the entire crust, and the depth and nature of the crust–mantle transition. Some inaccuracies were observed near the edges of the images due to the limited ray coverage inherent to the footprint of the survey geometry. Several aspects of the waveform tomography strategy were critical to the success of the acoustic method with realistic, synthetic, viscoelastic data: (i) the accuracy of the starting model from traveltime tomography, (ii) implementation in the frequency domain, (iii) the use of complex‐valued frequencies to effect time damping of the data residuals, (iv) the selection of a suitable subset of data and data frequencies, (v) progressive inversion of low‐ to high‐frequency components of the data, (vi) initial, pre‐inversion matching of the amplitudes between observed and modelled data, and (vii) sufficient preconditioning of both the data and the update images. Combined, these strategies were effectively equivalent to a multiscale approach that mitigated the non‐linearity of the seismic inverse problem. During the inversion we carried out repeated forward modelling to ensure our modelled waveforms matched the observed data as closely...