72nd EAGE Conference and Exhibition incorporating SPE EUROPEC 2010

The waveform inversion problem is inherently ill-posed. Traditionally, regularization terms are used to address this issue. For waveform inversion, where the model is expected to have many details reflecting the physical properties of the Earth, regularization and data fitting can work in opposite directions, slowing down convergence. In this paper, we constrain the velocity model with a model-space preconditioning scheme based on directional Laplacian filters. This preconditioning strategy preserves the details of the velocity model while smoothing the solution along known geological dips. The Laplacian filters have the property to smooth or kill local events according to a local dip field. By construction, these filters can be inverted and used in a preconditioned waveform-inversion scheme to yield geologically meaningful models. We illustrate on a 2-D synthetic example how preconditioning with non-stationary directional Laplacian filters outperforms traditional waveform inversion when sparse data are inverted for. We think that preconditioning could benefit waveform inversion of real data where (for instance) irregular geometry, coherent noise and lack of low frequencies are present.

Full Waveform inversion is a computer intensive process, especially for 3D seismic data. After a tremendous number of synthetic examples, finally real 3D data sets have been undertaken by the industry. As field data is dominated with P waves, one feasible approach is to use the acoustic approximation. The Full Waveform Inversion described above is acoustic and real data is more accurately described by an elastic model. It is common practice to apply acoustic inversion, especially for 3D data sets because the elastic modeling is prohibitively expensive. Although the long offsets may suffer from elastic effects our experiment shows that the velocity field obtained using acoustic inversion vs. elastic data are reasonable in spite of the difference between the modeling and mother Earth. Although the elastic propagation would provide a better match to the acquired data the cost is still prohibitive compared to the acoustic propagation especially if large 3Ds are under consideration.

The application of waveform inversion to land data is even more challenging than the marine case because of strong elastic effects such as groundroll, near surface attenuation, scattering due to rapid geological variations and topography. In this paper, rather than considering full elastic waveform inversion which maybe too difficult for standard geophone data, we consider the application of acoustic 2D full waveform inversion with a frequency domain code. At lower frequencies, the data are very noisy so we carried out a fixed number of iterations with a small band of low frequencies and then repeated this for increasingly larger bandwidth. We present the results from this procedure, which show the waveform inversion improved the continuity of the main reflectors.

Full waveform inversion is increasing being used in oil industry to quantify P and S-wave velocities of the sub-surface. So far, waveform inversion methods have been either 3D acoustic or 2D elastic. Furthermore, only low frequencies have been used because of the high computation cost. The application of elastic full waveform for 3D reservoir monitoring is still beyond the reach of present day computing technology. Here, we demonstrate the application an injected grid method, where the forward modelling for iterative elastic waveform inversion is performed in a small volume surrounding the reservoir, reducing the cost significant for full waveform. The application of algorithm is demonstrated on Marmousi II data set in a time-lapse mode.

Because of an increasing use of seismic inversion products within Maersk Oil, it was needed to perform a benchmarking exercise to identify seismic inversion/reservoir characterization contractor(s) of choice. To ensure robustness of the benchmark a highquality dataset was used (high SNR, good well coverage and well established rock physics model) from a clastic Palaecene reservoir showing a strong class II AVO response. The exercise included four contractors and Maersk’s internal inversion group, each asked to provide relative and absolute acoustic impedance, Poisson’s ratio volumes along with Bayesian seismo-facies prediction. The comparison was carried out by measuring correlations between inverted and measured relative acoustic impedance and relative Poisson’s ratio, combined with visual inspection of different sections, maps and along well locations in terms of predictability and continuity of events. We have found that the reliability in estimating acoustic impedance was similar for all the participants, and it was in fact the ability to estimate Poisson’s ratio that was the main differentiator. Other important aspects were significant differences in software runtimes, efficiency of workflows and processing/pre-conditioning capabilities. This benchmark has given us a sound overview of the abilities of the participating contractors.

4D pre-stack inversion is used in the industry to image reservoir changes due to production and injection, and to make reservoir management decisions in order to optimize hydrocarbon recovery. We present an innovative workflow to prepare, constrain and compute 4D pre-stack inversion attributes. Specific properties of the studied field (huge time-shifts due to gas coming out of solution, various turbiditic contexts) implied building a composite warping result, filtered using a 4D mask to build the initial monitor model for 4D inversion. The pre-stack 4D inversion workflow not only integrates seismic information, but also well information, used to discriminate sand from shale during the 4D mask building, and a 4D rock-physics model. Applied to simulated reservoir properties, the rock-physics model defines a range of relative density and velocity variations in which the inversion results can vary. Moreover, because water-bearing sands are hard to discriminate from shales in some of the field reservoirs using a cross-plot of P and S impedances, information from the reservoir grid was also introduced to help locating water-bearing sands in the 4D mask. Preliminary analyses of 4D inversion attributes show an improved image compared to previous 4D attributes.

The importance and impact of the way marginal (unconditional) distributions are considered in statistical rock physics and Bayesian classification of inverted seismic data was analyzed. Two scenarios were compared, one assuming perfect knowledge of the probability density functions and perfect and unbiased sampling; and another one where an unclassified group is included, with a given distribution in order to account for imperfections in the data, models, and estimation techniques. Using a dataset from a clastic Palaeocen field in the North Sea, it was shown that assuming the first scenario (perfectly known distributions) leads to probabilistic volumetric estimations of hydrocarbon saturated sands at least 3 times bigger than in the second case where an unclassified group is included. For the sake of further downstream engineering and facilities analyses, it is straightforward to realize the impact of such over-predictions in the estimation of P10, P50 and P90 volumes of hydrocarbons in place.

We develop a methodology for estimation of the point-spread function of prestack depth migrated data (PSDM). The parametrization of the point-spread function is given by a ray-based approach which incorporates the effects of wave propagation through the use of illumination vectors. The only unknown factor in the expression for the point-spread function is a 1D pulse. This pulse is estimated from reflection coefficients in wells, and co-located PSDM data using a least squares approach. The estimate of the point-spread function is the first step towards seismic inversion of PSDM data. In comparison to traditional wavelet estimation, the model using PSDM data has a larger range of validity, i.e. we are able to remove the assumption of a horizontally layered earth. In examples we show that our method is identical to the 1D convolution when the earth has a constant dip, but gives an improvement when this assumption is violated.

There has been considerable interest recently in data acquisition using simultaneous sources because of the enormous improvements in acquisition efficiency and source sampling that the method proffers. Realizing these improvements in practice requires an appropriate combination of survey design, acquisition technology and data processing capabilities. Given a suitably-designed survey, I show that simultaneous-source data can be separated effectively into equivalent datasets for each source. These datasets may then be processed using conventional techniques, which benefit naturally from any improvements in sampling. Several field datasets, employing a variety of acquisition geometries, illustrate the viability and limitations of this approach. The main limitation comes from ambiguities in the separation process, which can be resolved to a large degree by using source dithering techniques in combination with a separation algorithm that includes an effective sparseness constraint. The use of more than two sources simultaneously adds to the potential of the method and is shown to be viable provided the survey is designed appropriately.

This paper demonstrates a method of producing pre-stack gathers from blended acquisition data that are as noise-free as gathers from conventional acquisition. Filtering out the interference from the overlapping shots and stacking the data are fairly effective in attenuating the interference from blended shots. However high-quality separation of the interference from the pre-stack data would make the data more suitable for amplitude dependent analysis, such as that for amplitude-versus-offset, time-lapse measurements, and fracture detection. The method presented here produces seismic records in which the interference is attenuated to the point that it is well below the background noise. This high quality attenuation is achieved by using a sparse inversion process that solves a modified version of Berkhout's matrix system, allowing the source responses to be separated for high quality amplitude measurements. The success of this source separation step in the data processing of blended acquisition means we can produce results with a quality that is comparable to that of conventional acquisition. Furthermore, this method improves the data quality while retaining the lower cost and higher productivity of ISSTM acquisition when compared with conventional acquisition.