USING SEISMIC INVERSION AND GEOSTATISTICS TO ESTIMATE POROSITY: A WESTERN CANADIAN REEF EXAMPLE

Paper presented at the 1997 CSEG Convention JOHN PENDREL AND PAUL van RIEL Jason Geosystems, 700, 250 6th Ave. SW, Calgary, AB, Canada T2P 3H7 Tel: (403) 263-3340 Fax: (403) 266-7201

ABSTRACT We analyze data acquired over a Devonian reef to demonstrate how 3D seismic inversion and 3D geostatistical analysis can be used to estimate porosity and uncertainty. The result is a set of porosity simulations, all of which are equally probable. The simulations, computed as 3D volumes, are constrained to be in agreement with both log and seismic data and the geologic model for the reservoir. Summary maps can be used to view the average porosity in any layer or combination of layers. The combination of 3D geologic modeling and high resolution constrained sparse spike seismic inversion has proven to be a useful tool in the exploration for hydrocarbons (Pendrel and Van Riel, 1997). Not only do these methods account for the effects of wavelet sidelobes but also they admit the possibility of combining all available information into a single rock property product, an estimate of the impedance. An outstanding problem has been the question of uncertainty and subsequent confidence in the final result. This arises naturally in any scientific investigation but in particular here, because of the inherent non-uniqueness of the problem. We apply a 3D geostatistical analysis based on computed estimates of the probability density functions (pdfs) of the inputs and their spatial correlations (variograms) to estimate variability in the result. We use the constrained sparse spike inversion cube corresponding to the project data as a co-simulator. This is computed using all of the available information - seismic, interpreted horizons, geology and logs. Among the logs are porosities from neutron-density cross-plots. We determine cross-variograms between porosity and impedance which reflect both the table information and the spatial correlation of the impedance. Using the estimated pdfs and variograms, we do 3D sequential Gaussian co-simulation of the porosity logs with the inversion impedance cube as the secondary input. To further reduce uncertainty, we also undertake geostatistical inversion, wherein the co-simulation process is conditioned such that the 3D output models match the input seismic data. Several simulations are done and a set of possible porosity outcomes computed. The simulated outcomes are three-dimensional, allowing a full analysis of both lateral and vertical heterogeneity and variability. Summary information from the simulations can be determined by computing porosity means and standard deviations at each sample in the 3D grid and by extracting maps. We focus not on the mathematics of the method itself which have been described elsewhere (Journel and Huijbregts, 1978) but on its application to a data set.

INTRODUCTION Geology The geology is that of a Western Canadian Devonian reef. Figure 1 shows the structure of the Reef horizon and the locations of the wells used. The width of the main build-up is about 1 km. Surrounding the main reef is a porous apron which interfingers with tighter facies. A lateral salt is present which is dissolved close to the main build-up. There were several episodes of dissolution. Some were compensated by the basin material and others were not. This has had some influence in the small scale structure of the off-reef events. The off-reef basin facies are dolomitic and contain thin anhydrites which onlap against the reef. The considerable velocity contrast between the porous reef core and the high impedance off-reef facies has resulted in push-down of the reef platform and lower horizons. Other major reflectors are a

shale-carbonate transition above the reef (Carb) and a second deeper laterally continuous salt under the reef platform. Exploration issues centre around the distribution of apron porosity.

Figure 1: Structure of the Reef horizon and the locations of the vertical and horizontal wells. Geologic Model The structure of the 3D geologic model is defined by two pieces of information - the interpreted horizons and the model framework. The framework, in the form of a spreadsheet, describes the ordering of the horizons in space and time and their behaviour at faults. The horizons, which can include interpreted faults, provide structure information. Together, these form a blueprint for the model. Stratigraphy within the layers is specified as onlap, offlap, reef, channel, etc. The model is completed by populating it with geophysical information, usually input in the form of well logs. The logs are most commonly provided in depth and the horizons in time. It is therefore necessary to create a time-depth transformation if these two pieces of information are to be rationalized. Input sonic logs are integrated, hung on an input time datum and drift-corrected to tie the time horizons. Upon completion, the model exists in both time and depth and time-to-depth and depth-to-time transformations are possible. In addition to regarding the completed model as a useful end result, we use it in several other applications. It becomes a source of low frequencies for the sparse spike inversion. Since the low frequencies in the output inversion can not be controlled by the seismic, the model represents our most reliable information about them. Sparse spike inversion constraints are estimated at the wells and then interpolated between wells, guided by the structure of the model. In geostatistical modelling, stochastic simulation is performed within the model layers, while honouring the geologic relationships. Finally, summary maps measure average geophysical properties across the project, and do so in layers conformable with the horizons in the model. Constrained Sparse Spike Inversion The inversion is performed by a constrained sparse spike method as described by Pendrel and Van Riel (1997). First, wavelet analysis is accomplished by computing a filter which best shapes the well log reflection coefficients to the input seismic at the well locations. The phase of the input seismic data, which

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 2

can vary with frequency, is found from the inverse of this filter. We seek to model the input seismic data as the convolution of the seismic wavelet with a reflection coefficient series. Because the seismic wavelet is band-limited, the inversion problem is non-unique. There are many reflection coefficient series which when convolved with the seismic wavelet reproduce the input seismic to arbitrary accuracy. Thus, agreement of the sparse spike synthetic with the seismic becomes a necessary but not a sufficient condition in the solution of the inversion problem. To find the best geophysical and geological solution from the large number of available mathematical solutions, we must impose other conditions. These are provided by geophysical constraints which describe how the impedance can vary laterally away from the wells. A fairway of allowed impedances is defined, which limits the variability of inversion impedances and automatically rejects many good mathematical solutions which are geophysically or geologically not justifiable. The impedance fairway is conformable with the layering in the model and also has complete variability in time. These specifications force the inversion modeling problem to become non-linear, the solution to which is affected by an iterative technique. The number of iterations themselves, become a constraint, controlling the sparsity or blockiness of the inversion. Knowledge of the expected trace-to-trace impedance variations can also be provided to further limit the solutions. Since there is no guarantee in the mathematics that there will be agreement between the computed inversion and the impedance logs, quality control can be done by comparing them. Although, the constraints in sparse spike inversion will guarantee the existence of low frequencies in the inversion, we do not expect that the information in the lowest frequencies will be reliable. This is a consequence of the band-limited nature of the seismic wavelet. The lowest reliable frequency will be project dependent, depending upon the quality of the input seismic and the constraints applied to it. A merge frequency can be defined, below which, information in the final inversion is provided by the model. The model can be computed using logs from a single well or many, depending upon the circumstances. The former option will guarantee that no lateral varying information beyond that due to structure can come from the model while the latter will include the most available information about the project area. Above the merge frequency, information in the inversion comes from the constrained sparse spike output as described above. Geostatistical Analysis In geostatistical analysis, the statistics of the data are generated in the form of histograms and variograms. The sample histograms as measures of the relative frequency of occurrence of rock properties, are estimates of the corresponding probability density functions (pdfs). The histograms are modeled as Gaussian (or log Gaussian) distributions, which are in turn used in the simulation algorithm to transform the input data to a zero-mean, unit-standard deviation state. In this context, the modeled histograms are referred to as transforms. Variograms describe the spatial distribution of properties and convey the same information as the more familiar covariance function although they are defined to measure mis-correlation. Formally, the variograms, (k) are related to the covariances, (k) through (k) = (0) (k), where k is the lag

Variogram values plotted as a function of lag, are a minimum at zero lag where there is maximum correlation. At large lags, they approach a constant maximum value of mis-correlation corresponding to the power in the data. This maximum value is called the Sill. The lag at which the variogram value becomes 95% of the Sill is called the Range. In this, stationarity of both mean and variance is assumed. Departures from stationarity can affect both the variograms and the quality of the results. In this paper, variograms are estimated on a layer basis and respect the stratigraphy structure and faulting in the model. Problems in geostatistics centre about the optimum estimation of a primary input at grid points where its value is unknown. First, the transforms, variograms and the values at known neighbour grid points are used to estimate a local pdf at an unknown grid point. The unknown values are themselves estimated by randomly extracting a value from the estimated pdf. This is referred to as simulation. Many simulations can be done as required. Since the input data are usually 3D, the estimated variograms must also be 3D and the simulations must be done at grid points in a 3D volume. Furthermore, the grid points are chosen in random order to remove bias and preserve the variance in the input data.
Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997 Page 3

When the simulations are completed with the assistance of a secondary input, the process is referred to as co-simulation. Usually, the primary input is known at well control with high vertical resolution but low spatial resolution. The co-simulator data contains less vertical resolution but has higher spatial resolution, being known at many or all grid points. The goal of co-simulation is to distribute the primary data in three dimensions across the geologic model, constrained by the lateral variation provided by the secondary input. In this paper, we do co-simulation of porosity with inversion impedance as the secondary input. While the process described above preserves the spatial variance in the primary input data, there is no guarantee that any of the simulations are consistent with the input seismic data in an optimum sense. This situation can be rectified by computing synthetics after each simulation and comparing them to the input seismic, discarding simulations which result in poor matches. If impedance is not the primary input being simulated, then a table relating the primary to impedance must be provided. The selection criterion is controlled by a simulated annealing schedule. Early in the modeling, some poorer simulations are kept in order to ensure that the final simulation set is optimum in a global sense. We expect that the set of simulations produced using this geostatistical inversion will show a reduced standard deviation, since each is directed toward agreement with the seismic data. RESULTS Data The project data consisted of 100 lines, each with 90 CMPs. The bin size was 40 * 40 m. Time migration before stack had been done and care taken to preserve true amplitudes. Seventeen vertical wells, each with sonic and density logs were used. Computed porosity logs were available for many. Several horizontal wells have been drilled and were used for quality control. The key interpreted horizon was the Reef event which was picked along the top of the build-up until the salt was encountered whereupon the salt base was followed. Other key horizons were the shale carbonate interface above the reef, the reef platform and the tops of both salts. We checked that the well tops were picked consistently and that they were geophysical. By this we mean that they should be the well log counterparts of the interpreted horizons. Any necessary drift corrections were applied in the modeling step. EarthModel A solid geologic model with a 2 ms sample interval was constructed from the well logs, interpreted horizons and a framework describing the arrangement of the layers. The framework was particularly simple since there were no faults and layers were built up, one upon the other. A section through the 3D model is shown in Figure 2. The broad colours are meant to indicate the layer boundaries. The structure of the finer banding internal to each layer is indicative of the stratigraphy. In most layers, the stratigraphy is conformable with the bounding horizons. Exceptions are the Salt layer, which is conformable only with its upper boundary, and the reef interior which has been defined to respect both the reef structure and the observed push-down. Figure 3 is the same section (line 40 - see Figure 1) through the final impedance model. Lateral variations in impedance arise from both structural and stratigraphic effects. At CMPs where no log is present, nearby logs were stretched or squeezed on a layer-by-layer basis to account for structure and then subjected to a weighted average to arrive at the model impedance. A simple distanceweighted average sufficed although the weights could have been set arbitrarily as required. We used the model as a source of low frequencies in the final inversion. The reef mass is a low frequency effect and we will need these frequencies if the reef is to be imaged correctly.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 4

Figure 2: A stratigraphic section through the 3-D geologic model. The colours indicate the different layers defined by the interpreted horizons and the framework table. The banding within each layer describes the layer stratigraphy, which governs how lateral well log interpolation is done.

Figure 3: Line 40 through the 3D geologic model. The computed impedance logs have been interpolated according to the framework defined by the interpreted horizons.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 5

Constrained Sparse Spike Inversion Constraints were set within each geologic layer to limit the range of solutions to those which were geophysically admissible. These constraints, shown at well A in Figure 4, were assessed at each well and then interpolated across the project area according to the model horizons. Note that we do not expect higher impedances in the Reef layer than those observed in the log. However, we do want to include inversion solutions which speak of lower impedances since we expect to encounter porosity. The impedance fairway reflects these beliefs.

Figure 4: Fairway of allowed impedances at well A. The fairway width determines the range of admissible solutions to the sparse spike modeling within each geologic layer. Figure 5 is a section through the final inversion cube. The input seismic has been transformed to a layer domain. It now includes our best information about the low frequencies and the interpreted horizons. Relatively low regions of impedance can be seen inside the reef core and in the apron. We have measured average impedance variations by making maps. The reef apron event was copied both below and above itself, every sample for 20 ms. After re-building the model, it was then possible to measure the average impedance in any one or combination of one-sample-thick layers. Since all of the layers were conformable with the apron horizon we could measure the lateral variation in average apron impedance. Figure 6 is a map for a sub-apron layer 4 ms thick, starting at the Reef event. Porosity development is evident to the south below well P, to the west at well L and to the north and north-west. The good porosity near the reef core is in agreement with the horizontal well observations.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 6

Figure 5: Line 40 through the final inversion cube. The magenta colour approximately represents the limit of the best porosity.

Figure 6: Average impedance in a layer 4 ms thick immediately below the Reef horizon. Porosity development is evident to the north-west and near west and south of the reef core. Porosity logs computed from neutron-density crossplots were available at some of the wells. Corresponding average impedances were measured from the final inversion cube and the data plotted in Figure 7. Also shown is the best linear fit to impedance-porosity pairs from the Well G logs. The good agreement implies that valid porosities can be determined from the inversion. The average impedance in the layer immediately below the Reef horizon was converted to porosity and is shown in Figure 8.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 7

Figure 7: Average inversion impedances at the wells are plotted against average well log porosities (blue squares). An impedance-porosity function from well G logs is superimposed.

Figure 8: The average impedances in Figure 6 have been converted to porosities using the table in Figure 7. Geostatistical Analysis Histograms, variograms and cross-variograms were computed corresponding to all of the inputs to the geostatistical analysis. Figure 9 shows the histogram for impedance in the Reef layer. A Gaussian distribution has been modeled and provides an acceptable fit. The Gaussian model was used to transform the input impedances to zero mean and unit standard deviation. Similar transforms were computed for the porosity logs inputs. Since the porosity-impedance relation in Figure 7 was assumed, the porosity
Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997 Page 8

histogram was simply a re-scaled version of the impedance histogram. Three-dimensional variograms were computed for all the geostatistical inputs since the vertical and horizontal statistics are expected to be quite different. The cross-variogram between porosity and impedance is shown in Figure 10. Note the differences in the modeled ranges in the horizontal and vertical directions. In this analysis no transverse anisotropy was assumed and the horizontal variogram was made independent of azimuth. There are few samples in the vertical component of the sample variogram because the Reef layer is thin. However, a small range is indicated and it has been so modeled. The cross-variogram in the figure contains both information about the porosity-impedance relation in Figure 7 and the spatial variability of impedance and porosity.

In v e r s io n I m p e d a n c e H i s t o g r a m
0 .1 0

P r o b a b il it y

0 .0 8
S ta n d a rd

0 .0 6 0 .0 4 0 .0 2 0 .0 0
M ean

D e v ia t io n

1 .0

1 .2

1 .4

1 .6

1 .8

2 .0

In v e r s io n Im p e d a n c e x 1 0 7

Figure 9: A histogram of the impedance distribution in the Reef layer, to which a Gaussian probability density function has been modeled.

Cross-Variogram of Porosity with Impedance

Cross-Variogram Value

-0.2 -0.4 -0.6 -0.8 -1.0 0 200 400 600 800 1000

Horizontal Vertical

1200

1400

Lag Distance (m)

Figure 10: The 3D cross-variogram between porosity and impedance. The modeled vertical range is much smaller than the horizontal range, reflecting the difference in resolution in these two directions. The ranges in the two horizontal directions have been modeled to be the same.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 9

Twenty co-simulations were done with log porosity as the primary input and constrained sparse spike impedance as the co-simulator. Four of these co-simulations are shown in Figure 11. The simulations, while generally consistent, do show differences depending upon location. For example, it is evident that there is much more variability to the left of the figures in the apron.

Figure 11: Line 44 from four different porosity simulations. No geostatistical inversion was done. Considerable variability is evident in the left (south-central) area of the figures. A mean porosity cube was computed from the 20 simulations and is shown in Figure 12. A standard deviation cube was also found and is shown in Figure 13. Together, they give a 3D representation of the most probable porosity distribution and the expected uncertainty in this estimate. The standard deviation is larger farther away from the main build-up. Here, the Reef layer is thinner, there are fewer data points, and less well control. The highest porosities are to be found close to and within the main build-up. This implies a lateral non-stationarity and potentially unrealistic simulations. As expected, the standard deviation is small close to well control. The co-simulations were re-run, constrained by geostatistical inversion. In the inversion algorithm, a synthetic was determined after each simulated sample was computed. The synthetic was compared to the seismic and a simulated annealing schedule used to determine whether the simulation was to be retained. The result was a set of 20 simulations, each of which, while different, was in agreement with the input seismic data to an arbitrary degree of accuracy. Line 44 through the mean and standard deviation geostatistical inversion cubes are shown in Figure 14 and 15. Note that the standard deviations are generally smaller for geostatistical inversion. Apparently, many of those previous simulations were not in agreement with the seismic. This is especially true away from the main build-up where the gains have been the greatest. Summary maps of the mean and standard deviation in the 4 ms layer below the reef apron were constructed as described above. (Figures 16 and 17). Porosity development in the windward north-eastern direction is evident. An average porosity of 12 % is indicated with a standard deviation of up to 2 %. Four horizontal wells have recently been drilled for apron porosity.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 10

Figure 12: Line 44 through the Mean cube derived from the average of 20 simulation cubes. No geostatistical inversion has been done. Also shown are porosity logs from five nearby wells.

Figure 13: Line 44 through the standard deviation cube corresponding to the 20 simulation cubes computed without geostatistical inversion. Five nearby well tracks have been indicated. Note that the standard deviation is a minimum near these sources of known information. The standard deviation is highest in the apron region where the well control is poorest

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 11

Figure 14: Line 44 through the Mean cube derived from the average of 20 simulation cubes with geostatistical inversion. Also shown are porosity logs from five nearby wells

Figure 15: Line 44 through the standard deviation cube corresponding to the 20 simulation cubes computed with geostatistical inversion. The standard deviation has been reduced generally and particularly in the apron region. Shown in Figure 16, they are wells U, Y, Z, to the north-east and X to the south. Well U first detected apron porosity in the windward north-east direction. Well Y has been the most successful of these wells to date and well Z caught the edge of porosity close to the main build-up. To the south, well X was a
Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997 Page 12

disappointment. After encountering initial porosity, the facies became tight. All of these observations are in very good agreement with Figure 16.

Figure 16: Average mean co-simulated porosity in the 4ms layer below the apron. There is an enhanced porosity region in the windward direction to the north-east.

Figure 17: Average standard deviation in the 4 ms layer just below the Reef apron from the 20 co-simulations with geostatistical inversion.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 13

While the most desirable domain in which to make an interpretation is depth, an important consideration is the choice of datum. Accurate depths are available at the wells. In addition to this sparse information, there are less accurate but spatially dense, interpreted times at each CMP. This is another example of a problem which can be solved with geostatistics. Depth tops can be co-kriged with interpreted times as the secondary input. We performed collocated cokriging of the Carb tops to obtain a continuous horizon in depth. Then, we used this result and the geologic model to convert the volume and horizons below the Carb to depth. The Reef horizon, converted to depth is shown in Figure 18. By copying the Reef depth horizon repeatedly below itself, average porosity and -h maps can be computed for depth layers conformable with it. Figure 19 is a -h map from the average geostatistical inversion porosity cube for the 10 m layer immediately below the apron. Since 10 m corresponds approximately to the time interval in Figure 16, the two figures are very similar.

Figure 18: The depth-converted Reef horizon. Carb depth tops were cokriged with Carb interpreted times as the secondary input. The volume and horizons belwo the Carb were then converted to depth using the geologic model. CONCLUSIONS Quantification of uncertainties in geophysical measurements such as impedance and porosity is necessary for complete risk analyses. This is an important consideration in both exploration and reservoir characterization situations. We have shown that the methods of geostatistics can be used to define limits to the expected values of porosities in a Devonian reef. The analysis must be completed using 3D statistics if the multi-dimensional properties of the geophysical parameters are to be correctly modeled. Essential and unique to the proposed method to reduce uncertainty is that the 3D output result fully honours the input well logs, geologic model and seismic data. Finally, the estimation of porosity in a thin layer underneath the reef apron was determined to be in good agreement with recent observations from horizontal drilling.

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 14

ACKNOWLEDGEMENTS The authors gratefully acknowledge Husky Oil Ltd. and Mobil Oil Canada for their generous donation of these data. REFERENCES Journel, A.G., Huijbregts, C.J. (1978): Geostatistical Reservoir Characterization Constrained by 3D Seismic Data, Ext. Abs., 58th EAGE meeting, Amsterdam Pendrel, J.V., Van Riel P. (1997): Methodology For Seismic Inversion, A Western Canadian Reef Example, CSEG Recorder, 22, #5

Using Seismic Inversion and Geostatistics to Estimate Porosity April, 1997

Page 15