Geophysical Journal International

Transcription

1 Geophysical Journal International Geophys. J. Int. (2012) 189, doi: /j X x Cross-well seismic waveform tomography for monitoring CO 2 injection: a case study from the Ketzin Site, Germany Fengjiao Zhang, 1 Christopher Juhlin, 1 Calin Cosma, 2 Ari Tryggvason 1 and R. Gerhard Pratt 3 1 Department of Earth Sciences, Uppsala University, Villavagen 16, Uppsala, SE-76236, Sweden. fengjiao.zhang@geo.uu.se 2 Vibrometric Oy, Taipaleentie 127, Perttula, Finland 3 Department of Earth Sciences, University of Western Ontario, London, Ontario, N6A 5B7, Canada Accepted 2012 January 9. Received 2012 January 9; in original form 2010 November 3 1 INTRODUCTION The CO2SINK project is the first onshore European CO 2 injection experiment. The CO2SINK site is located at Ketzin, west of Berlin, Germany (Fig. 1). Three wells have been drilled at the site, one serving as the CO 2 injection well and the other two as observation wells (Fig. 2). The injection started on 2008 June 30, and as of 2011 July 31, tons of CO 2 have been injected in the underground ( The CO 2 is injected into a saline aquifer in the heterogeneous Upper Triassic Stuttgart Formation. which is located at approximately m depth in the injection well. The Stuttgart Formation is about 80 m thick and is comprised of siltstones and sandstones interbedded with mudstones. The saline sandstone aquifers vary in thickness from 1 to 30 m (Norden et al. 2010). The caprock for the CO 2 reservoir is comprised of the Weser and Arnstadt Formations, containing almost 210 m of mudstone and evaporate. A m-thick high-velocity anhydrite layer (K2) is present at the top of the Weser formation. As CO 2 replaces saline water in saturated reservoir sandstones an up to 21 per cent P-wave velocity reduction may occur (Xue & Lei 2006; Xue et al. 2006). This velocity change can potentially be SUMMARY Geological storage of CO 2 is one means of mitigating the effects of continued burning of fossil fuels for power generation. An important component in the storage concept is the monitoring of the CO 2 distribution at depth. Seismic methods can play a significant role in this monitoring, in particular cross-well methods are of interest due to their high resolution. For these purposes, a series of cross-well seismic surveys were acquired within the framework of the CO2SINK project at Ketzin, Germany, at various stages of an injection test. We study here the potential of applying cross-well seismic waveform tomography to monitor the CO 2 injection process. First, we test the method on synthetic data having a similar geometry to that of the real data. After successful application on the synthetic data, we test the method on the real data acquired at the Ketzin Site. Traveltime tomography images of the real data show no observable differences between the surveys. However, seismic waveform tomography difference images show significant differences. A number of these differences are artefacts that can probably be attributed to inconsistent receiver coupling between the different surveys. However, near the injection horizon, below the caprock, a velocity decrease is present that is consistent with that expected from the injection process. Key words: Inverse theory; Downhole methods; Controlled source seismology; Seismic tomography; Computational seismology. used to monitor CO 2 in sandstone aquifers using seismic tomography. A number of previous applications of using tomography as a tool for monitoring CO 2 injection have been reported. Lazaratos & Marion (1997) used seismic traveltime tomography to monitor reservoir changes due to CO 2 injection at Chervron s McElroy Field in West Texas. Daley et al. (2008) applied time-lapse traveltime tomography to cross-well data from the Frio CO 2 injection site and found a velocity decrease in the sandstone aquifer of about 500 ms 1 due to the presence of CO 2. Saito et al. (2006, 2008) applied time-lapse cross-well seismic traveltime tomography for monitoring CO 2 injection in an onshore aquifer at Nagaoka, Japan, and their results show an observed decrease in velocity of 3 per cent. Spetzler et al. (2008) applied time-lapse difference traveltime seismic tomography to data from the same Nagaoka injection field. Their results showed that the injected CO 2 migrates upwards in the porous sandstone reservoir and that the observed velocity perturbation due to CO 2 injection is about 18 per cent. Onishi et al. (2009) applied the difference analysis with data normalization (DADN) method to time-lapse tomography data, also from the Nagaoka test field, and their results show an about 9 per cent decrease in velocity. Gosselet & Singh (2007, 2008) and Singh & Queisser (2010) applied GJI Seismology C 2012 The Authors 629

2 630 F. Zhang et al. Figure 1. (a) Location of the Ketzin pilot site, Ketzin, west of Berlin, Germany, (b) perspective air photo of the injection site and locations of wells KTZI-201, KTZI-200 and KTZI-202 (photo used in Fig. 1b is from Figure 2. A 3-D sketch of the geometry of the cross-well surveys at the Ketzin Site. full waveform inversion in the time-lapse mode for monitoring CO 2 sequestration at the Sleipner field. They compared elastic models of the subsurface computed through 2-D full waveform inversion and their results showed that it is possible to use full waveform inversion to quantify the amount of CO 2 injected. Ravaut et al. (2009) also applied waveform tomography to 2-D synthetic data based on the V p model at Sleipner and tried to better characterize the thickness and the V p velocities of the CO 2 -saturated sand layers. Hogan et al. (2009) made some investigations on synthetic data and their results suggested that waveform tomography deserves serious consideration as a primary analysis tool for time-lapse seismic analysis. At the Ketzin Site, Cosma et al. (2009) applied traveltime tomography to the time-lapse cross-well data acquired at the site to monitor the CO 2 injection. Their results indicated that conventional P-wave cross-well traveltime tomography is not able to map the CO 2 distribution at the time of the surveys since no significant first arrival traveltime variations were observed. However, some changes in the reservoir were revealed by traveltime difference tomography. In this paper we study the potential of time-lapse seismic waveform tomography (full waveform inversion) to monitor CO 2 injection at the Ketzin Site. First, we apply the method to synthetic data using an acquisition geometry similar to the real data. After successfully applying the method to the synthetic data, we then apply it to the real data and compare our results with those using only traveltime tomography. Finally, we discuss our results in more general terms regarding CO 2 monitoring using seismic waveform tomography. 2 METHOD Seismic waveform tomography is a powerful and advanced tool to reconstruct the underground velocity field in high resolution using seismic data. Formal methods for waveform inversion were originally developed about 25 yr ago (Tarantola 1984; Mora 1988). These methods have been further enhanced by many others (Pratt 1990; Pratt & Worthington 1990; Song et al. 1995; Pratt et al. 1998, 2005; Zhou & Greenhalgh 2003; Ravaut et al. 2004; Sirgue & Pratt 2004; Brenders & Pratt 2007a,b; Brossier et al. 2009; Sourbier et al. 2009a,b; Vireux & Operto 2009). There are a number of examples of successful applications of the method on field data, in particular from cross-well geometries (Pratt & Shipp 1999; Pratt et al. 2005; Rao et al. 2006; Wang & Rao 2006; Barnes et al. 2008). This, along with the indication that waveform tomography provides

3 images with higher resolution than ray-based methods using first arrival traveltimes, should make the method ideal for time-lapse studies where the affected volumes may be below the resolution limits of traveltime tomography. In this paper we use the frequency domain waveform tomography method first introduced by Pratt and colleagues in the late 1990s (Pratt et al. 1998). They showed that frequency domain waveform tomography could produce similar results to time domain approaches by using just a few frequencies from the data set and inverting for them in the frequency domain. In the frequency domain, the velocity model is updated by first inverting the low frequency components, and then gradually moving to higher frequency components to increase the resolution. Pratt & Symes (2002) also pointed out that traveltime tomography and waveform tomography are complementary to each other. Traveltime tomography needs to be performed first in real data applications since seismic waveform tomography requires a good starting model for the method to be successful. This is because of the highly non-linear nature of seismic waveform tomography. Details of the frequency domain waveform tomography method are described by Pratt et al. (1998). Here we summarize the method for the sake of completeness. The acoustic wave equation in the frequency domain can be written as S (m; ω) u (m,ω) = f (ω), (1) where S is the differencing matrix, u the wavefield and f the source vector. Eq. (1) can be rewritten as u = S 1 f. (2) In the inversion problem we want to minimize the L2 norm of the data residuals, so we define an objective function as E(m) = 1 2 δd T δd, (3) where T denotes the matrix transpose and represents the complex conjugate. The term δd is the residual wavefield and m is the model parameter vector. To minimize the objective function we use the gradient method to update the model parameter vector iteratively by m (i+1) = m i y i m E i, (4) where i and γ are the iteration number and step length, respectively. The gradient vector can be evaluated by taking partial derivatives of eq. (3) with model parameters m, ( ) [ E S m E (i) = = Re {u t t m m (i) i ] } v. (5) In this equation u is the scaled forward modelled wavefield and v = ( S 1) t d. (6) is the backpropagated residual wavefield. The inversion method may be easily divided into three steps for each iteration (Wang & Rao 2006). First, calculate the forward synthetic wavefield u based on a given initial model (starting model). Second, backpropagate the weighted data residual v to evaluate the gradient vector. Finally, estimate the model update. An important consideration when applying a 2-D waveform inversion algorithm to real data is the difference in the impulse response between 2-D (the modelling dimension) and 3-D (the acquisition dimension) media. This difference is handled by multiplying the extracted source function by 1/sqrt(iω), where ω is angular frequency, to compensate for this difference (Williamson & Pratt 1995). Waveform tomography for CO 2 monitoring 631 Table 1. Modelling parameters. Grid size Sample Source Record Model size (m) rate wavelet length z x = dz = ms First derivative 60 ms dx = 0.5 Gauss wavelet 3 SYNTHETIC DATA TEST 3.1 Synthetic cross-well seismic data Before applying seismic waveform tomography on the real crosswell seismic data set, we first carried out a synthetic test to verify if the method is suitable for the Ketzin data set. To make the tests more realistic, the synthetic seismic data sets were generated using the same geometry as the real data sets acquired at the Ketzin Site. The synthetic data sets were generated using an elastic 2-D finite difference code in Seismic Unix (Juhlin 1995). The modelling parameters are shown in Table 1. The velocity models used in the synthetic tests were based on geological information, laboratory results and well log data from the Ketzin Site (e.g. Kazemeini et al. 2010; Ivanova et al. 2012). For the modelling, the storage zone was set at m depth with a 5 15-m-thick plume and a reduction in velocity of 500 m s 1 in the zone containing CO 2.A velocity reduction of 500 m s 1 was chosen since this is close to what is observed in laboratory experiments on Ketzin core from the reservoir zone (Ivanova et al. 2012). The source was activated from 452 to 739 m at 1 m intervals along the KTZI-200 well (Fig. 2). The receivers were placed from 464 to 726 m at 1 m intervals along borehole KTZI-202, with 36 receivers per source location. The velocity models before and after CO 2 injections are shown in Fig. 3 with corresponding synthetic shot gathers. 3.2 Traveltime tomography and waveform tomography applied to the synthetic data sets After generating the synthetic data sets, we applied both traveltime tomography and waveform tomography to the data sets to compare the results from the two methods. For the traveltime tomography we used a code based on local earthquake tomography (Benz et al. 1996; Tryggvason et al. 2002; Bergman et al. 2004) and later modified to handle controlled source surveys (e.g. Yordkayhun et al. 2009). For seismic waveform tomography we used a code based on the frequency domain method (Pratt et al. 1998). To overcome the non-linearity of the waveform tomography methods we used the traveltime tomography result as a starting model for the waveform inversion. Before performing waveform tomography, some pre-processing steps were applied to the data sets. First, the data sets were filtered ( Hz) to be comparable with the real data sets. Then the model grid spacing was reduced to 0.5 m to make the computation stable at high frequencies. Finally, the data sets were tapered with a 5 ms time window after the picked first arrivals for the first run of the inversion and then tapered with a 10 ms time window to include more data for the later iterations of the inversion. After pre-processing of the data sets, waveform tomography was performed. First, the source signature was solved for using the linear approach introduced by Pratt (1999), starting from 300 Hz up to 1300 Hz at 50 Hz intervals. Then the velocity model was inverted for from 300 to 1200 Hz at 50 Hz intervals with 3 neighbouring frequencies as a group. We started with the 5 ms windowed data set and then used the results from the 5 ms windowed data set as the starting model for the 10 ms windowed data set. Fig. 4 shows a

4 632 F. Zhang et al. Figure 3. Velocity models and synthetic elastic data sets generated using the similar geometry as the real data set acquired at Ketzin. (a) Velocity models before and after CO 2 injection (the red arrows point out the position of the injections), (b) example of synthetic shot gathers before and after CO 2 injection, shot at depth 635 m (red star on the left-hand side of the models in subpart a), receivers from depth m (red line on the right-hand side of the model in subpart a). The red squares on the shot gathers show the picked traveltimes.

5 Waveform tomography for CO 2 monitoring 633 Figure 4. Comparisons between the traveltime tomography results and the final waveform tomography results on the synthetic data. (a) Traveltime tomography result for before CO2 injection, (b) traveltime tomography result for after CO2 injection, (c) waveform tomography result for before CO2 injection, (d) waveform tomography result for after CO2 injection, (e) velocity difference for traveltime tomography results (before injection after injection), (f) velocity difference for waveform tomography results (before injection after injection), (g) True model velocity difference (before injection after injection, the red arrows point out the position of the velocity reductions).

6 634 F. Zhang et al. comparison between the traveltime tomography results and the final waveform tomography results. In the traveltime tomography results, a velocity change caused by the CO 2 injection is clearly observed, but the magnitude and boundary of the anomaly do not match the true model as well as in the waveform tomography result. In particular, the small anomaly just above 650 m next to the receiver well (located at a distance of 100 m) cannot be positively identified in the traveltime tomography results (compare Fig. 4e with Fig. 4f). Tests with less smoothing applied in the traveltime tomography (allowing more short wavelength features into the model) did not provide a better match than that shown in Fig. 4. Our results thus show that waveform tomography may be more effective than traveltime tomography in recovering small velocity changes in the velocity model for these synthetic data sets using identical source and receiver geometries as for the real data set. Therefore, we were encouraged to test the seismic waveform tomography method on the real cross-well data sets acquired at the Ketzin Site to monitor CO 2 injection. 4 REAL DATA APPLICATION 4.1 Real data acquisition and processing A 3-D sketch of the geometry of the cross-well surveys at Ketzin is shown in Fig. 2. The distance between the source well KTZI-200 and the receiver well KTZI-202 is about 100 m. The injection well KTZI-201 is located 50 m away from the source well KTZI-200 and 120 m from the receiver well. As in the synthetic tests, the source was activated from 452 to 739 m at 1 m intervals along the KTZI-200 well. The receivers were placed from 464 to 726 m at 1 m intervals along borehole KTZI-202, with a maximum of 36 receivers per source location. A time-distributed VIBSIST piezoelectric borehole source (Cosma & Enescu 2001) and a 12-level hydrophone chain were used. The main technical specifications of the TC-12 hydrophones and their sensitivity to incoming wavefront angle are shown in Fig. 5(a). Fig. 5(c) shows the VIBSIST-SPH source, which is based on the Swept Impact Seismic Technique Figure 5. Hydrophone receiver, TC-12 and piezoelectric SPH-54 borehole source used for cross-well data acquisition at Ketzin. (a) Hydrophone receiver, TC-12 and key parameters of it, (b) amplitude response of a TC12 hydrophone for a calibrated source [laboratory measurements were made at 10 intervals in the xy- (horizontal) and xz- (vertical) planes], (c) piezoelectric SPH-54 boreholes source and its key parameters, (d) radiation pattern of a radial source in a vertical plane containing the borehole.

7 (SIST), and its main technical specifications. Signals are produced by applying electric pulses to a stack of piezoelectric crystals. The mechanical energy that is generated is conveyed to the rock by a system of wedges actuated by an electric motor. Each pulse generates about 4 8 J of seismic energy in a frequency band of Hz. The time-lapse cross-well seismic measurements were carried out during 2008 and 2009 to monitor the CO 2 injection. The baseline survey was acquired in 2008 May before CO 2 injected had started. The first repeat survey was performed in 2008 July after CO 2 had been detected in the closest observation well (KTZI-200 in Fig. 2), but not in the more distant well, KTZI-202. The second repeat survey was performed in 2008 August, again prior to arrival of CO 2 in the more distant observation well (KTZI-202 in Fig. 2). A small repeat survey was acquired in 2009 July. In this third repeat survey the number of hydrophone levels was reduced to four since gaseous CO 2 had filled the receiver well KTZI-202 above 650 m. About 650 tons of CO 2 had been injected at the time of first survey and about 1750 tons at the time of the second repeat. The main processing steps for the real data sets involved the following: (1) bandpass filtering of the data in the frequency range from 300 to 5000 Hz, and (2) suppression of tube waves in both the receiver and source domains. Fig. 6(a) shows a comparison between a raw and processed source gather from the baseline data set. Fig. 6(b) shows an example of processed source gathers from the baseline, the first repeat and the second repeat data sets. The source was at 630 m depth and the receivers span the interval m, thus covering the CO 2 injection reservoir zone. Fig. 7 shows the signal-to-noise ratio (S/R) as a function of source depth with different frequency bands. The S/R was estimated by calculating the rms amplitude in a 10 ms window after the first arrival and dividing these values by the rms amplitude in a 10 ms window before the first arrival for all traces. These results were then averaged for each shot location to give the average as a function of source depth. First arrival times show very little differences in the three source gathers as shown in Fig. 6(b). However, later arriving phases show more variability (red arrows in Fig. 6b). In particular, the approximately constant traveltime phase between 35 and 40 ms appears more delayed in the 2008 August survey. Based on the source gathers in Fig. 6(b) and similar ones at nearby levels, it is clear that traveltime tomography using solely first arrivals should not show any large velocity changes between the two wells. However, the delay in the later arriving subhorizontal phase indicates that seismic waveform tomography may reveal some differences. 4.2 Traveltime tomography results After data processing, we applied traveltime tomography to the baseline and the two repeat data sets using the picked first break times. We used the same traveltime tomography code as applied to the synthetic data. The starting velocity model for the traveltime tomography was derived by assuming straight rays between the two boreholes at equal depths. We simply divided the distance between the two wells by the picked first arrival times to generate a horizontally layered starting model. The model size was m in the x-direction and m in the z-direction with 1 1 m cells for Waveform tomography for CO 2 monitoring 635 both directions. The traveltime residuals were reduced to ms after six iterations, a reduction of more than 70 per cent compared to the starting value of ms. After six iterations the residuals showed only a minor decrease and additional iterations were just adding inversion noise into the results, therefore we stopped the inversion at this point. Fig. 8 shows the traveltime tomography results for the baseline, first and second repeat data sets. There are no significant differences between the baseline and the two repeats over the reservoir interval at about m. This is consistent with the seismograms shown in Fig. 6(b) where there are no significant changes in the first arrival times, but where some delays in the secondary arrivals are present. The latter observation indicates that waveform inversion may allow the velocity reduction due to the CO 2 injection to be observed for the present data set. Velocity differences at m are associated with the fast anhydrite layer and these are discussed later in the paper along with results from the waveform tomography. 4.3 Waveform tomography Data pre-processing Additional pre-processing steps that needed to be preformed on the real data before waveform tomography could be applied included trace editing, amplitude corrections, the 2-D to 3-D data correction and time windowing. Noisy or abnormal traces whose waveforms were not consistent with the other traces in the same shot gather were removed. Such traces may have a detrimental effect on the results of the inversion. Amplitude corrections are done to handle the problem that forward model data have amplitude differences compared with real data. This step tries to match the rms amplitude of the real data to the forward modelled data. The difference in the impulse response between 2-D (the modelling dimension) and 3-D (the acquisition dimension) media is handled by multiplying the extracted source function by 1/sqrt(iω). Time windowing was used to eliminate noise prior to the first arrival and to exclude later arriving shear wave energy from the data sets. This forces the inversion to match the first arriving energy, which contains important information on the low and intermediate wavenumbers in the model in the initial stages of the inversion (Pratt et al. 2005). A 6 ms window was applied to the data sets, starting from 1 ms before the picked first arrival. In Fig. 6(b) the blue lines in the figure indicate the 6 ms time window used in the inversion. A relevant question in the inversion of the real data is the dependence of amplitude on the incidence angle to the sources and receivers. Fig. 5 shows calibration measurements carried out at high frequency (20 khz) and in open water. In our case the maximum angle of incidence in the vertical plane is about 13.5, which is quite small (about 85 per cent of the amplitude compared with 0 for the hydrophone and source). Azimuthal variations are negligible according to the calibration measurements. Even though the angle of incidence in the vertical plane is small, we tested applying a linear correction of the amplitude to the data according to the source and receiver angles. The result indicated that the velocity differences between before and after the correction were relatively small (about 50 m s 1 ) compared with the velocity differences we are looking for (about m s 1 ). In addition, the wave will refract at the borehole wall, further reducing the incidence angle. For these reasons, we decided not to apply these amplitude corrections to the data.

8 636 F. Zhang et al. Figure 6. (a) Comparison between a raw (left panel) and processed (right panel) source gather from the baseline data set. (b) Example source gathers from the injection depth (630 m) after processing, red squares indicate the picked traveltime, blue lines indicate the time window used in the waveform tomography (time window is 1 ms before the first arrivals to the blue line) and red arrows indicate where the delay of the secondary arrivals is interpreted (blue arrows point out shot depth) Starting model An adequate good starting model is needed for waveform tomography to be successful. The starting model must be able to describe the time domain data sets to within a half-cycle at the lowest usable frequency to avoid matching the wrong cycle of the waveforms (Pratt et al. 2005). In our real data case the dominant frequency range is from 300 to 1300 Hz, thereby lacking low-frequency information. This makes the waveform inversion results strongly dependent on the starting model. Here, we used the traveltime tomography result from the baseline data set as the starting model for the waveform tomography. This same starting model was used for the baseline and repeat data sets (Fig. 9a). As a check on the suitability of the starting model we generated shot gathers based on it and compared C 2012 The Authors, GJI, 189, C 2012 RAS Geophysical Journal International

9 Waveform tomography for CO 2 monitoring 637 Figure 7. Signal-to-noise ratio as a function of source depth for different frequency bands. them with the real data. Figs 9(b) and (c) show an example of a synthetic shot gather generated from the starting model and the corresponding shot gather from the real baseline data set. The two shot gathers were filtered to the same frequency range, between 100 and 1000 Hz. The first breaks of the synthetic shot gather generated from the starting model match the arrival times of the first breaks in the real data set quite well Inversion scheme The first step in waveform tomography is to invert for the source signatures using the initial starting model. The frequencies used for the source signature inversion were from 100 to 1000 Hz at 20 Hz intervals. This wide range ensures that all the frequency components in the data set are included during the inversion. During the inversion

10 638 F. Zhang et al. Figure 8. Traveltime tomography results for the baseline (2008 May), first repeat (2008 July) and second repeat (2008 August) data sets, and the velocity differences between them (a) baseline traveltime tomography result, (b) first repeat traveltime tomography result, (c) second repeat tomography result, (d) velocity difference between baseline and first repeat (May July), (e) velocity difference between baseline and second repeat (May August), (f) velocity difference between first repeat and second repeat (July August).

11 Waveform tomography for CO 2 monitoring 639 Figure 9. (a) Starting model (from the traveltime tomography) for the waveform tomography for all the data sets. (b) An example of a real shot gather. (c) The corresponding synthetic shot gather generated from the starting model, red squares indicate the picked first break times and the blue lines indicate the time window used in the inversion. for the velocity field the initial source signature results were used and only the amplitudes of them were updated. Then we used the updated velocity model to get the new source signatures, which were then used for the next inversion stage. After the source signatures were estimated the waveform tomography inversion procedure for the velocity field was initiated. The first step was to choose which frequency components to invert for. The real data sets contain useful frequencies in the range from 300 to 1000 Hz and higher. We started from the lowest frequency component in the data set at 300 Hz in an attempt to overcome cycle-skip problems and continued to 1000 Hz in 20 Hz steps. Then we used three neighbouring frequencies as a group simultaneously in the inversion. The selected 36 frequency components were divided into 12 groups. The result of each lower frequency group was used as the starting model for the next group and for each group we did three iterations. We used a 5 ms window after the first arrival as our data window. During the inversion for the velocity field we filtered the gradient along the wells to remove undesirable large values. After each step we checked the results by generating synthetic seismograms based on the resulting model Waveform tomography result Fig. 10 shows a comparison between the sonic log velocities, the traveltime tomography result and the baseline waveform tomography result at the source and receiver wells. The traveltime and waveform tomography results exhibit a comparable fit with the sonic logs. However, the waveform tomography result matches the sonic logs slightly better than the traveltime tomography result in some places (marked by red arrows) for example in the CO 2 injection zone at around m depth. Fig. 11 shows the final waveform tomography results for the velocity fields and the differences between the different data sets. The most pronounced changes in velocity are in the depth range m and m. These apparent changes in velocity are discussed in the following section. In any inversion method, including waveform tomography, the most important question is how well the final results can predict the observed data. Fig. 12 shows a comparison between real shot gathers and synthetic shot gathers generated from the final waveform tomography results. The synthetic shot gathers show some similarity with the real shot gathers, not only with the first arrival times. To quantify the results, we also compared the rms differences of the waveforms between the real data and the generated synthetic data based on the starting model and the final inversion results (Table 2). These show that the synthetic data generated by the final inversion results match the real data better than that generated by starting model. These observations indicate that we have had some degree of success with our inversion. 5 DISCUSSIONS The prominent change in the velocity field at m (Fig. 11) is associated with the anhydrite layer, located well above the reservoir zone. Strong changes in velocity over this interval were also observed in the traveltime inversion results (Fig. 8). These changes cannot be due to the injection of CO 2. Instead, we attribute it to inversion artefacts due to noise from either different hydrophone coupling conditions between surveys or to the sources/sensors not being located at identical depths during the different surveys (Marelli et al. 2010; Maurer et al. 2010). By coupling, we mean here the rock-hydrophone response. The casing was not cemented in this interval and conditions in the annulus may have changed between the

12 640 F. Zhang et al. Figure 10. Velocities from the sonic logging (blue line) compared with traveltime tomography (black line) and waveform tomography (red line) results in the wells KTZI-202 and KTZI-200 for the baseline data set (2008 May). surveys. CO 2 may have migrated up the outside of annulus in between the surveys as indicated by logging data (Ivanova et al. 2012). Given the high velocity of the anhydrite layer, small differences in station locations or coupling could result in significant differences in the inverted velocity fields. Evidence for poor receiver coupling is found in plots of the average autocorrelation of the data sets in both the receiver domain and source domain (Fig. 13). At each depth the average autocorrelation for either the receiver gather or the source gather is shown. The autocorrelation for the source gathers is fairly consistent, but the receiver gathers show significant variation, especially around the anhydrite layer at 550 m. The variation around the anhydrite layer appears to be increasing in extent with time of the survey, suggesting deteriorating near-borehole conditions in this area. Variations in the average receiver autocorrelation function are also observed at and m, but the vertical extent of these variations do not appear to be increasing with time, suggesting

13 Waveform tomography for CO 2 monitoring 641 Figure 11. The final waveform tomography result for the baseline, first repeat, second repeat and the differences between them. (a) Baseline waveform tomography result, (b) first repeat waveform tomography result, (c) second repeat waveform tomography result, (d) velocity difference between baseline and first repeat (baseline first repeat), (e) velocity difference between baseline and second repeat (baseline second repeat) and (f) velocity difference between first repeat and second repeat (first repeat second repeat).

14 642 F. Zhang et al. Figure 12. Example shot gathers from the real data and the corresponding synthetic shot gathers generated from the final waveform tomography results (red squares indicate the picked first arrival traveltime, blue lines indicate the time window used in the waveform tomography). (a) P-wave velocity log at about reservoir depth, red and blue arrows point the position of selected example shots and the same colour lines point out the receiver arrays of the same shot. (b) An example shot gather comparison between real data sets and forward modelled data sets generated from the final inversion results (the left-hand column is real data and right-hand column is synthetic data). (c) Another example shot gather comparison between the real data sets and the forward modelled data sets generated from the final inversion results (the left-hand column is real data and right-hand column is synthetic data). Table 2. rms differences of the waveforms between synthetic data generated by starting models, synthetic data generated by final inversion results and real data. rms difference of the waveforms between synthetic data generated by starting model and real data rms difference of the waveforms between synthetic data generated by final inversion results and real data Baseline First repeat line Second repeat line (2008 May) (2008 July) (2008 August) that borehole conditions are not changing at this depth. However, they may be the cause of some of the anomalies at these depths near the boreholes (Fig. 14). Near the injection horizon ( m), velocity differences between the baseline and repeat surveys vary from 700 to 700 m s 1 (Fig. 14). The repeat surveys show a decrease in velocity of about 700ms 1 compared to the baseline at about the injection depth. This velocity difference is similar to what may be expected due to the presence of CO 2 based on laboratory experiments (Ivanova et al. 2012). The main positive velocity changes between the baseline (2008 May) and the first repeat (2008 July) are at about m between the two wells. The velocity differences between the baseline and second repeat (2008 August) show a continued decrease in velocity at about m (marked in Fig. 14 with ellipse). We interpret the shallow ( m in Fig. 14 ) and deeper (about 655 m in Fig. 14) anomalies to be artefacts due to the poorer

15 Waveform tomography for CO2 monitoring 643 coupling of the geophones at this depth as discussed earlier, while the positive anomaly between 630 and 635 m (marked in Fig. 14 with ellipse) may be due to CO2. It is important to note that direct geochemical monitoring in the wells showed CO2 arriving in KTZI-200 on 2008 July 21 and in KTZI-202 first on 2009 March 28 (Martens et al. 2011), that is long after the second repeat cross-well survey was acquired and long after that predicted from fluid flow modelling (Kempka et al. 2010). This late arrival is probably due to the heterogeneity of the Stuttgart Formation as also evidenced in 3-D surface seismic data at the site (Kazemeini et al. 2009). Due to the heterogeneity of the formation, it is possible that CO2 could be channelling into a zone between the two wells during injection. The seismic waveform tomography images do not provide an unambiguous set of images to be interpreted. However, the images are roughly consistent with observations. The main anomaly is found near the top of the reservoir in the injection horizon (at about m) and is of the correct sign, indicating lower velocities after injection. This main low-velocity anomaly increases C 2012 The Authors, GJI, 189, C 2012 RAS Geophysical Journal International in size between the two repeat surveys, as would be expected. Note that since the same starting model was used for the inversion of the baseline and the two repeats that the two repeats then represent two independent measurements of the velocity changes. In both cases, the anomaly just below 630 m appears. This is further evidence that it is real and that its location suggests that it is due to CO2. The anomaly does not extend all the way to the KTZI-202 well, which is also consistent with observations. Anomalies above and below the injection horizon are interpreted as due to the inconsistent coupling of the receivers at about and as shown in Fig. 13. The most reliable time-lapse change is that indicated by the subhorizontal velocity decrease between 630 and 635 m at the base of the caprock, suggesting that this anomaly is due to CO2 injection, but the geometry of it should not be overinterpreted. The actual distribution of CO2 is probably more spread out and the inversion is localizing the anomaly to a smaller area. Regardless, the seismic waveform inversion is providing more information than it was possible to obtain using only traveltime inversion. Figure 13. Autocorrelations of the data sets in both the receiver and source domains. The traces represent the average correlation of all traces in each gather. (a) Autocorrelations of the data sets in the receiver domain and (b) autocorrelations of the data sets in the source domain.

16 644 F. Zhang et al. Figure 14. (a) The velocity differences between the baseline and the first repeat compared with the geological log information and (b) the velocity differences between the baseline and the second repeat compared with the geological log information. Red arrow and vertical lines marked the reservoir formations. Finally, we note that the more horizontal velocities, as determined from the tomography, are consistently higher than the vertical sonic log velocities in the caprock (rock above ca.630 m in Fig.10),In the Stuttgart Formation (rock below ca. 630 in Fig. 10) the difference is much less. This indicates that the reservoir rock is nearly isotropic whereas the caprock has about 5 10 per cent anisotropy in the P- wave velocity. This anisotropy will have an influence on our velocity models that we have not taken into account. However, since we are

17 looking for differences in the velocity models, the anisotropy should affect all three surveys equally and any differences in velocities due to changes in physical conditions should still be observable. 6 CONCLUSIONS In this paper we presented a case study of time-lapse cross-well seismic waveform tomography for monitoring CO 2 injection. In the synthetic data test the waveform tomography method accurately reconstructs the true velocity field and shows a better capability to recover small changes in an underground velocity model than traveltime tomography. This indicates that waveform tomography could be a useful tool for monitoring the velocity changes caused by CO 2 injection at the CO2SINK site at Ketzin and at other locations. Classical P-wave traveltime tomography was not capable of detecting any velocity differences caused by CO 2 injection in the real data. Application of seismic waveform tomography to the real data showed some zones of decreased velocity between the two observation wells. Some of these zones may be artefacts due to poor receiver coupling, but a subhorizontal anomaly at the base of the caprock may be due to CO 2 injection. The location of the anomaly, the observed velocity change and the fact that it appears in the same location on both repeats are evidence for it being due to the injection of CO 2. However, the anomaly should not be overinterpreted. No CO 2 had arrived at the far well (KTZI202) at the time of the repeat surveys. The lack of a continuous channel of CO 2 between the wells makes interpretation of the seismic waveform tomography results more difficult. ACKNOWLEDGMENTS The CO2SINK project was funded by the European Commission, the German Federal Ministry of Education and Research, the German Federal Ministry of Economics and Technology as well as research institutes and industry, under project no We thank the CO2SINK consortium for the complementary data set and publication permission. Critical reviews by Hansruedi Maurer and one anonymous reviewer helped to significantly improve this paper. Thanks to Alireza Malehmir in helping setting up the input data for the waveform inversion and to Niklas Juhojuntti, Can Yang and Nicoleta Enescu for helpful discussions. REFERENCES Barnes, C., Charara, M. & Tsuchiya, T., Feasibility study for an anisotropic full waveform inversion of cross-well data, Geophys. Prospect., 56, Benz, H., Chouet, B. Dawson, P., Lahr, J., Page, R. & Hole, J., Three dimensional P-and S-wave velocity structure of Redoubt Volcano, Alaska, J. geophys. Res., 101, Bergman, B., Tryggvason, A. & Juhlin, C., High-resolution seismic traveltime tomography incorporating static corrections applied to a till covered bedrock environment, Geophysics, 69, Brenders, A.J. & Pratt, R.G., 2007a. Efficient waveform tomography for lithospheric imaging: implications for realistic 2D acquisition geometries and low frequency data, Geophys. J. Int., 168, Brenders, A.J. & Pratt, R.G., 2007b. Full waveform tomography for lithospheric imaging: results from a blind test in a realistic crustal model, Geophys. J. Int., 168, Brossier, R., Operto, S. & Virieux, J., Seismic imaging of complex structures by 2D elastic frequency-domain full-waveform inversion, Geophysics, 74, Waveform tomography for CO 2 monitoring 645 Cosma, C. & Enescu, N., Characterization of fractured rock in the vicinity of tunnels by the swept impact seismic technique, Int. J. Rock Mech. Min. Sci., 38, Cosma, C., Enescu, N., Cosma M. & the CO2SINK Team, Borehole seismic monitoring at CO2SINK site, in Proceedings of the 9th International Workshop on the Application of Geophysics to Rock Engineering, Hong Kong, China. Daley, T.M., Myer, L.R., Peterson, J.E., Majer, E.L. & Hoversten, G.M., Time-lapse crosswell seismic and VSP monitoring of injected CO 2 in a brine aquifer, Environ. Geol., 54, Gosselet, A. & Singh, S.C., Monotoring of CO 2 sequestration at sleipner using full waveform inversion in time-lapse mode, Am. geophys. Un., Fall Meeting 2007, abstract # H12D-07. Gosselet, A. & Singh, S.C., D Full waveform inversion in time-lapse mode: CO 2 quantification at Sleipner, in 70th EAGE Conference & Exhibition, EAGE Expanded Abstract WO11, 69 73, Hogan, C., Hedlin, K., Margrave, G. & Lamoureux, M., Feasibility testing of time-lapse seismic monitoring with full waveform tomography, in Proceedings of the Frontiers+Innovation-2009 CSPG CSEG CWLS Convention, Calgary, Alberta, Canada. Ivanova, A., Kashubin, A., Juhojuntti, N., Kummerow, J., Juhlin, C. & Lüth, S., D seismic data analysis at Ketzin, Germany, and CO 2 volumetric estimation using petrophysical data, core measurements and well logging: a case study, Geophys. Prospect., in press. Juhlin C., 1995, Finite difference elastic wave propagation in 2-D heterogeneous transversely isotropic media, Geophys. Prospect., 43, Kazemeini, H., Juhlin, C., Zinck-Jørgensen, K. & Norden, B., Application of the continuous wavelet transform on seismic data for mapping of channel deposits and gas detection at the CO2SINK site, Ketzin, Germany, Geophys. Prospect., 57, Kazemeini H., Juhlin C. & Fomel S., Monitoring CO 2 response on surface seismic data: a rock physics and seismic modeling feasibility study at the CO 2 sequestration site, Ketzin, German, J. appl. Geophys., 71, Kempka, T., Kühn, M., Class H., Frykman P., Kopp, A., Nielsen, C.M. & Probst, P., Modelling of CO 2 arrival time at Ketzin: part I, Int. J. Greenhouse Gas Control, 4(6), Lazaratos, S.K. & Marion, B.P., Crosswell seismic imaging of reservoir changes caused by CO 2 injection, Leading Edge, 16, Marelli, S., Manukyan, E., Maurer, H.R., Greenhalgh, S.A. & Green, A.G., Appraisal of waveform repeatability and reliability for crosshole and hole-to-tunnel seismic monitoring of radioactive waste repositories. Geophysics, 75(5), Q21 Q34. Martens, S., Liebscher, A., Möller, F., Würdemann, H., Schilling, F., Kühn, M. & Ketzin Group, Progress report on the first European on-shore CO 2 storage site at Ketzin (Germany): second year of injection, Energy Procedia, 4, Maurer, H.R., Greenhalgh, S.A., Marelli, S., Manukyan, E. & Green, A.G., Systematic errors in waveform inversions caused by variable receiver coupling. SEG Expanded Abstr, 29, 2757, doi: / Mora, P., Elastic wavefield inversion of reflection and transmission data, Geophysics, 53, Norden, B., Forster, A., Vu-Hoang, D., Marcelis, F., Springer, N., & LeNir, I., Lithological and Petrophysical Core-Log Interpretation in the CO2SINK, the European CO 2 Onshore Research Storage and Verification Project, SPE Reservoir Eval. Engineer., 13(2), Onishi, K., Ueyama, T., Matsuoka, T., Nobuoka, D., Saito, H., Azuma, H. & Xue, Z.Q., Application of crosswell seismic tomography using difference analysis with data normalization to monitor CO 2 flooding in an aquifer, Int. J. Greenhouse Gas Control, 3, Pratt, R.G., Frequency domain elastic wave modelling by finite differences: a tool for crosshole seismic imaging, Geophysics, 55, Pratt, R.G., Seismic waveform inversion in the frequency domain, part 1: theory and verification in a physical scale model, Geophysics, 64,

18 646 F. Zhang et al. Pratt, R.G. & Shipp, R.M., Seismic waveform inversion in the frequency domain, part 2: fault delineation in sediments using crosshole data, Geophysics, 64, Pratt, R. G., & Symes, W., Semblance and differential semblance optimization for waveform tomography: a frequency domain implementation, in Proceedings of Sub-Basalt Imaging Conference, Expanded Abstracts, pp Pratt, R.G. & Worthington, M.H., Inverse theory applied to multisource crosshole tomography I: acoustic wave-equation method, Geophys. Prospect., 38, Pratt, R.G., Shin, C. & Hicks, G.J., Gauss-Newton and full Newton methods in frequency-space seismic waveform inversion, Geophys. J. Int., 133, Pratt, R.G., Hou, F., Bauer, K. & Weber, M., Waveform tomography images of velocity and inelastic attenuation from the Mallik 2002 crosshole seismic surveys, in Scientific Results from the Mallik 2002 Gas Hydrate Production Research Well Program, Mackenzie Delta, Northwest Territories, Canada, Vol. 585, p. 14, eds Dallimore, S.R. & Collett, T.S., Geological Survey of Canada Bulletin. Rao, Y., Wang, Y.H. & Morgan, J.V., Crosshole seismic waveform tomography II. Resolution analysis, Geophys J Int, 166, Ravaut, C., Operto, S., Improta, L., Virieux, J., Herrero, A. & Dell Aversana, P., Multiscale imaging of complex structures from multifold wideaperture seismic data by ferquency-domain full-waveform tomography: application to a thrust belt. Geophys. J. Int., 159, Ravaut, C., Alerini, M., Ghaderi, A. & Dillen, M., Full-waveform inversion for CO 2 monitoring: application to the Sleipner field, in Proceedings of SEG Summer Research Workshop, Banff, Canada. Saito, H., Nobuoka, D., Azuma, H., Xue, Z. & Tanase, D., Time-lapse crosswell seismic tomography for monitoring injected CO 2 in an onshore aquifer, Nagaoka, Japan, Explor. Geophys., 37, Saito, H., Nobuoka, D., Azuma, H., Tanase, D. & Xue, Z., Time-lapse crosswell seismic tomography for monitoring CO 2 geological sequestration at Nagaoka pilot project site, J. Mining Mater. Process. Inst. Japan, 124(1), Singh, S.C, & Queisser, M., Quantitative seismic monitoring of CO 2 at Sleipner using 2D full-waveform inversion in the time-lapse mode, in 72nd EAGE Conference & Exhibition, EAGE Expanded Abstract K013, Sirgue, L., & Pratt, R.G., Efficient waveform inversion and imaging: a strategy for selecting temporal frequencies, Geophysics, 69, Song, Z., Williamson, P. & Pratt, R.G., Frequency-domain acousticwave modeling and inversion of crosshole data, Part 2: inversion method, synthetic experiments and real-data results, Geophysics, 60, Sourbier, F., Operto, S., Virieux, J., Amestoy, P., & L Excellent, J.Y., 2009a. FWT2D: a massively parallel program for frequency-domain fullwaveform tomography of wide-aperture seismic data Part 1: Algorithm Comput. Geosci., 35, Sourbier, F., Operto, S., Virieux, J., Amestoy, P., & L Excellent, J.Y., 2009b. FWT2D: a massively parallel program for frequency-domain full-waveform tomography of wide-aperture seismic data Part 2: Numerical examples and scalability analysis, Comput. Geosci., 35, Spetzler, J., Zue, Z., Saito, H. & Nishizawa, O., Case story: time-lapse seismic crosswell monitoring of CO 2 injected in an onshore sandstone aquifer, Geophys. J. Int., 172(1), Tarantola, A., Inversion of seismic reflection data in the acoustic approximation, Geophysics, 49, Tryggvason, A., Rognvaldsson, S.T. & Flovenz, O.G., Threedimensional imaging of the P- and S-wave velocity structure and earthquake locations beneath Southwest Iceland, Geophys. J. Int., 151, Virieux J. & Operto, S., An overview of full-waveform inversion in exploration geophysics, Geophysics, 74, WCC127 WCC152. Wang, Y.H. & Rao, Y., Crosshole seismic waveform tomography - I. Strategy for real data application, Geophys. J. Int., 166, Williamson, P.R. & Pratt, R.G., A critical-review of acoustic-wave modeling procedures in 2.5 dimensions, Geophysics, 60, Xue, Z. & Lei, X., Laboratory study of CO 2 migration in watersaturated anisotropic sandstone, based on P-wave velocity imaging, Explor. Geophys., 37, Xue, Z., Tanase, D. & Watanabe, J., Estimation of CO 2 saturation from time-lapse CO 2 well logging in an onshore Aquifer, Nagaoka, Japan, Explor. Geophys., 37, Yordkayhun, S., Tryggvason, A., Norden, B., Juhlin, C. & Bergman, B., D seismic traveltime tomography imaging of the shallow subsurface at the CO2SINK project site, Ketzin, Germany, Geophysics, 74, G1 G15. Zhou, B. & Greenhalgh, S.A Crosshole seismic inversion with normalized full-waveform amplitude data, Geophysics, 68,