Professional Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Keywords: This study integrates 3D seismic attributes and well log data to determine Vsand and reservoir property distri-
Seismic inversion bution of the Oligo-Miocene Asmari Formation in the western Dezfol Embayment, SW Iran. The rocks consist of
Multi-attribute complex geology with siliciclastic and carbonate lithology. Hence, it creates the need for a precise interwell
Neural network estimation of lithology and reservoir properties. For this purpose, first the acoustic impedance was obtained by a
Reservoir property model-based inversion algorithm. Then, the acoustic impedance attribute and other sample-based seismic attri-
Mansuri oil field
butes were integrated with sand volume and petrophysical data by using multiple attribute regression and neural
Asmari reservoir
networks in order to predict Vsand and reservoir property. In the next stage, cross-validation was used to estimate
the reliability of the derived multi-attribute transforms. Based on the results of neural networks, the highest cross-
correlation was observed between seismic attributes and the observed target logs at seven wells in the study area.
After validation, and making a comparison between different available techniques for sand volume, effective
porosity and water saturation estimation, multiple regression transform and neural network were used for the first
two and latter, respectively. The derived sand volume and reservoir property maps for the Asmari reservoir
indicated that high-porous and high-sand volume parts were laterally more continuous in the central and east part
of the area under study. In addition, high-porosity zones were more related to high sand volume parts. Based on
the result of interpretation and the relationship between core and acoustic impedance, variations in acoustic
impedance were related to variations in geological characteristics of Asmari reservoir in the field. Therefore,
seismic inversion as a powerful tool can facilitate the detailed studies of sedimentary facies and lithology in the
reservoir which contribute to understand the subsurface reservoirs heterogeneities and drilling strategy of future
drilling campaigns in the study area.
* Corresponding author.
E-mail address: kadkhodaie_ali@tabrizu.ac.ir (A. Kadkhodaie).
https://doi.org/10.1016/j.petrol.2017.11.060
Received 8 October 2016; Received in revised form 14 October 2017; Accepted 22 November 2017
Available online 24 November 2017
0920-4105/© 2017 Elsevier B.V. All rights reserved.
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 2. Cenozoic stratigraphic correlation chart of the Iranian Sector of the Zagros Basin, adopted from James and Wynd (1965).
260
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 5. Time horizons of top Asmari formation (left) and top Pabdeh formation (right).
Formation span in age from the Early Oligocene (Rupelian) to the Early depositional environment occurred as a result of sea level falling so that
Miocene (Burdigalian) (Stocklin, 1968). The formation is mainly the unconsolidated medium to fine-grained sands could be precipitated
composed of fractured dolomites with limestones, and locally the Ahwaz as a lowstand systems tracts at some depths (Research Institute of Pe-
sandstone Member. The intervals of Kazhdumi Formation are regarded as troleum Industry (RIPI), unpublished). It is worth noting that the
the source rock while Gachsaran Formation is the cap rock. above-mentioned unconsolidated channel sands have the best reservoir
Based on the core and petrographical studies of the Asmari samples, quality in the Asmari reservoir.
with an emphasis on diverse ichnofossil assemblages, the depositional Based on lithology and well log evaluation, Asmari reservoir was
environment of the Asmari Formation changed several times in Mansuri divided into 8 main zones in this field which mostly include limestone
field from shallow marine to coastal plain (from offshore/lower shore- and sandstone. Here, sandstones encompass a majority of the hydrocar-
face through lagoonal to inter/supratidal flat) in a series of depositional bons of reservoir due to high porosity and low water saturation.
cycles. A shallow marine depositional system is interpreted as the sedi- In this reservoir, most of the production come from zone 1, which
mentary environment for the mixed carbonate-siliciclastic deposits of the mainly consists of limestone, along with zone 2 and 3 dominat by
Asmari Formation. The mixed carbonate-siliciclastic deposits indicating a sandstone with some interbeds of carbonate and shale (EDCO Kish In-
low input of clastic sediment to the shoreface occurs mostly within arid ternational Company, unpublished).
settings. The low clastic input may result in minimal fluvial input or low
relief in the shoreline and source area (Zonneveld et al., 2001). 3. Method and data collection
The fluvial input to the Asmari shoreface was characterized by
washover fan deposits and the abundance of quartzose, very-fineto fine- In this study, some of the conventional logs including sonic (DT), bulk
grained sand in lower shoreface, the presence of burrows in intertidal density (RHOB) and calculated logs of sand volume (Vsand), water satu-
deposits and coarse-grained sand infilling cracks within foreshore and ration (SW) and effective porosity (PHIE) were used. The latter log was
upper shoreface. computed using neutron-density crossplots in seven wells drilled in
It seems that the Asmari succession was deposited in six deepening- Asmari reservoir of the mansuri field. In order to calculate the effective
upward parasequence sets. Each parasequence is started by lagoonal porosity correctly, the shale volume was estimated and the porosity of
carbonate facies and continues with fine to very fine-grained shoreface the shale was eliminated from all computations.
sandstones. The influx of fine to medium-grained sands into the Asmari An integrated workflow was regarded as the method in which a well-
261
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 6. A sample of well to seismic tie for one of the wells in the Mansuri field showing the synthetic seismic generated from wavelet and well data (blue trace) and sampled seismic data
close to the well (red) and a correlation co-efficient of 69%. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 7. An arbitrary line from seismic data which passes through four well locations.
to-seismic tie was established to identify the key stratigraphic surfaces results of AI, horizon maps, and seismic attributes were computed from
which were interpreted to generate horizon maps. The generated hori- 3D post stack data by using multi-attribute regression analysis and neural
zons maps, along with necessary log properties in thirteen wells, were networks, which were used for predicting sand volume, effective porosity
used in acoustic impedance (AI) inversion. The inversion procedure in- and water saturation in the field under study.
volves well to seismic calibration, wavelet estimation, low-frequency After outlining the method for deriving the acoustic impedance and
model estimation and model based inversion for a seismic dataset of predicting reservoir properties, the results were described and inter-
Mansuri oilfield (Huuse and Feary, 2005; Kumar et al., 2016). Then, the preted based on diagenetic features and depositional setting of the
262
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 8. Arbitrary line from model based inversion impedance cube, passing through well locations. Inserted color at the well locations is acoustic impedance. Lateral and vertical changes of
Asmari reservoirh is clear after inversion. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Fig. 9. (a) Crossplot of actual and predicted volume of sand using multi attribute regression. Data points from analysis zone of each well are shown in one color. (b) Crossplot of predicted
effective porosity versus actual effective porosity. The correlation coefficient is 0.76. (c) Cross plot of actual and predicted water saturation (SW). (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)
Asmari reservoir in the field. called model-based inversion (e.g. Russell and Hampson, 1991).
As shown in Appendix A, model-based inversions are the generalized
4. Acoustic impedance inversion linear algorithm which uses an iterative forward modeling and com-
parison procedure (Simm et al., 2014). In this method, an initial model of
Seismic inversion is a sophisticated process of transforming the the acoustic impedance is first used and this model is perturbed until
seismic reflection data into a quantitative rock property, descriptive of computing the resulting synthetic trace by applying equations (A-1) and
the reservoir (see Appendix A). The most popular inversion technique for (A-2) which match original seismic data within some tolerance level. For
invert seismic data and create cube of full band acoustic impedance is each trace, a synthetic seismogram is calculated by using the extracted
263
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Table 1 synthetic seismogram was created by convolving the reflectivity with the
Multi-attribute list from a stepwise regression for prediction of sand volume, effective appropriate wavelet. The appropriate statistical wavelet for each well
porosity and water saturation in Asmari reservoir of the Mansuri field.
was acquired by using the extraction and selection of different wavelets,
Target Final Attribute Training Validation which constructed a best correlation between synthetic and real traces. It
Error Error
is worth noting that a good correlation between synthetic traces derived
1 Sand Volume Inversion Result 0.209811 0.217231 from well logs and original seismic traces at each well location comes
2 Sand Volume Filter 5/10–15/20 0.196126 0.206486 from a good approximation of seismic wavelet. (Kadkhodaie-Ilkhchi
3 Sand Volume Time 0.188816 0.201080
4 Sand Volume X-Coordinate 0.182628 0.194286
et al., 2013). After obtaining wavelet for each well and doing calibration
5 Sand Volume Integrate 0.178872 0.193005 process for each of the thirteen wells, all calibrated wells were used
6 Sand Volume Derivative Instantaneous 0.175361 0.190671 together to extract one average wavelet, which was used for the inver-
Amplitude sion. Figs. 4 and 5 illustrate the average wavelet and the picked time
7 Sand Volume Cosine Instantaneous Phase 0.172457 0.192846
horizons, respectively. A sample of a well-to-seismic tie, for one of the
8 Sand Volume Dominant Frequency 0.171007 0.200654
wells in the Mansuri field, is shown in Fig. 6 where the correlation be-
1 Porosity Log (Inversion Result) 0.054156 0.055292 tween synthetic seismogram (blue) and composite trace (red), at the well
2 Porosity Filter 5/10–15/20 0.050487 0.053074
3 Porosity Amplitude Envelope 0.047960 0.051193
location, is 0.69.
4 Porosity Time 0.046266 0.049623
5 Porosity Integrated Absolute 0.044453 0.048566 6. Initial model building
Amplitude
6 Porosity Cosine Instantaneous Phase 0.042924 0.046778
7 Porosity Amplitude Weighted 0.042170 0.047468 Seismic data are naturally band-limited due to the lack of high and
Frequency low frequencies. A low-frequency component is highly important for
8 Porosity Second Derivative 0.041735 0.047715 describing the geology and reservoir properties, which need to be ob-
1 Water Time 0.171589 0.174140 tained from other data (Latimer et al., 2000). In model-based inversion,
Saturation low-frequency data is usually generated by interpolation based on log
2 Water Log (Inversion Result) 0.152315 0.158315 data, horizon interpretation, and processing velocities. Hence, in the
Saturation
present study, it was supplied from the initial model based on the
3 Water Filter 5/10–15/20 0.142008 0.151253
Saturation interpolation of impedance logs from the thirteen wells along with the
4 Water Cosine Instantaneous Phase 0.138149 0.149652 interpreted seismic horizons (top of Asmari and top of Pabdeh forma-
Saturation tions). One arbitrary seismic line passing through the well locations is
5 Water Dominant Frequency 0.135606 0.150742 shown in Fig. 7.
Saturation
6 Water Average Frequency 0.132607 0.151145
The model based inversion was run with the appropriate wavelet and
Saturation the background impedance model, resulting in an impedance model
7 Water Integrated Absolute 0.131071 0.161834 which needs to be closely compared with the actual impedance in the
Saturation Amplitude well as a QC step. Fig. 8 displays a final impedance section for one of the
8 Water Amplitude Weighted Cosine 0.129353 0.159839
studied seismic arbitrary line (violet displays tight or high AI reservoir
Saturation Phase
intervals and green indicates porous or low AI intervals).
wavelet and the initial guess impedance. Then, the inverted impedance 7. Sand volume and reservoir property estimation
data is gradually modified during each iteration until matching the
resulting synthetic trace by the real trace within a predefined tolerance Some studies indicated a positive relationship between seismic at-
limit (Russell, 1988). Fig. 3 illustrates the work flow of the seismic tributes and reservoir parameters, sedimentary facies and geological
inversion for 3D-seismic of the Mansuri field. structures. In recent years, different researchers studied the linear and
The inversion procedure is summarized as below: non-linear relationships between seismic attributes and reservoir prop-
erties by geostatistics, neural networks, pattern recognition, artificial
Synthetic seismogram generation(Seismic to well tie and wavelet intelligence, and the genetic algorithm (Bosch et al., 2010; Iturraran-
extraction) Viveros and Parra, 2014; Kadkhodaie-Ilkhchi et al., 2014; Pearson and
Seismic Horizon picking Hart, 1999; Perez-Mu~ noz et al., 2013; Raeesi et al., 2012; Snedden, 2013;
Low-frequency model building(create initial model based on acoustic Soubotcheva and Stewart, 2006; Yao and Journel, 2000).
impedance) In the present study, seismic attributes derived from the 3D post stack
Model-based inversion data and acoustic impedance as an external attribute were used to find
the best relationship for estimating Vsand and reservoir property. In order
A well-to-seismic tie was established through a synthetic seismogram to find a reliable relation between seismic attributes, as input data, and
in order to identify two horizons of Asmari and top of Pabdeh formations. petrophysical parameters, as target data, multi-attribute regression
Then, these horizons were picked for using in initial model building for analysis and neural network algorithm were used to predict the Vsand and
the model-based inversion. reservoir property.
5. Well correlation and wavelet extraction 8. Application of single attribute regression analysis for sand
volume and reservoir property prediction
In the first stage of seismic inversion, seismic and well logs data were
imported to the inversion project, and well logs for each of the thirteen Assuming a linear relationship between target data and the seismic
wells were converted to two-way travel time using check shot. Then, the attribute and using the least squares method to determine the relation-
product of sonic and density logs yielded an acoustic impedance log at ship is regarded as the simplest procedure to establish the relationship
each well location. This impedance log was used to derive a log of between the data and the seismic attribute.
reflectivity. In the next procedure, the reflectivity was converted from The relation between target logs and different seismic attributes in-
depth to time by available check shot data for each of the well. Then, a dicates an acoustic impedance from seismic inversion having the highest
264
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 10. Plot of average error versus number of attributes to determine optimal number of attributes to use in prediction processes. Validation plot for all wells showing that the optimal
number of attributes to predict volume sand (a), effective porosity (b) and water saturation (c) are six, six and four, respectively. The black curve indicates the error using all wells, whereas
the red curve shows the error plot when a well is removed. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
Table 2
Shows the summary of sand volume, effective porosity and water saturation prediction by single and multi-attribute regression analysis, and neural network for studied field. In this table,
Significant improvement is achieved in application of multi attribute and neural networks rather than single attribute regression.
degree of correlation and the lowest degree of error related to effective water saturation more accurately in the Asmari reservoir.
porosity and sand volume. Further, time attribute has the highest cor-
relation with water saturation. Although a single attribute transform 9. Application of multi-attribute regression analysis for sand
presents good results, it does not yield a realistic value for 100% each of volume and reservoir property prediction
target logs. Importantly, acoustic impedance and time are not the only
single seismic attribute to predict the target logs of the reservoir. In fact, A multi-regression analysis is a simple and practical method to inte-
many attributes could be combined to derive the best multi-linear grate the information of numerous seismic attributes for predicting a
regression relationship between seismic data as extracted attributes target parameter. The objective of this method is to provide a group of
and the petrophysical properties. Here, a multi-linear regression was seismic attributes comprising better information than each single attri-
used to estimate target logs such as sand volume, effective porosity and butes can offer. The basic theory of multi-attribute regression for
265
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Table 3
Characteristics of Asmari facies in differentiated clusters based on acoustic impedance (AI).
Low-AI facies min 4750 0 0 Fine to medium/coarse-grained sandstones which is mainly clean, unconsolidated to
Cluster1 mean 6650 0.23 0.67 semi-consolidated, oil-saturated and appear massive without visual sedimentary structures.
max 7926 0.33 1
Medium-AI facies min 7930 0 0 composed of two distinctive lithofacies; 1: fine-grained sandstones with dolomitic cement
Cluster2 mean 9007 0.12 0.34 and argillaceous matrix and 2: dolomitic sandy skeletal packstone to grainstone with
max 10610 0.29 1 interparticle, biomoldic, vuggy and intercrystalline porosity.
High-AI facies min 10610 0 0 mud-supported facies and claystones.
Cluster3 mean 12210 0.09 0.23
max 14640 0.25 0.9
Fig. 11. Thin section photomicrographs and core photographs of cluster lithologies from the Asmari reservoir, Mansuri field. a) burrowed fine to medium-grained dolomitic sandstone
with interparticle porosity; Plan polarized light. b) burrowed fine-grained argillaceous sandstone; Plan polarized light. c) skeletal grainstone with interparticle and moldic porosity;
interparticle pore spaces have been occupied by calcite cement; Cross polarized light. d) burrowed dolomitic sandy mudstone with pervasive anhydrite plugging; Cross polarized light.
266
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 12. Sand volume prediction showing a sand volume distribution along transect B–B’. Sand volume gradually increase towards the north of field. Also, high-sandy zones are laterally
more continuous in the north part.
modeling reservoir parameters using different attributes is presented in predicting sand volume, effective porosity and water saturation, a com-
Appendix B. bination of six, six and four attributes, respectively, are statistically sig-
In application of multi-attribute regression, convolutional operator nificant, since the validation error starts to increase after this points.
and stepwise regression are two important issues. As shown in Appendix Then, these selected attributes were used to create the estimates of the
B, The optimal value of operator is obtained by minimizing the prediction target logs at the seven wells ties. The result of applying the derived
error between the predicted and the actual target logs. multi-attribute transform by using the selected attributes for sand vol-
On the other hand, stepwise regression is used for selecting an ume, effective porosity and water saturation indicate a correlation co-
optimal number of seismic attributes, in which the choice of predictive efficient of 0.77, 0.76 and 70, with an average prediction training error of
variables is conducted by an automatic procedure. (Draper et al., 1966; 0.175, 0.042 and 0.13, respectively (Table 2).
Hampson et al., 2001; Khoshdel and Riahi, 2011; Ogiesoba, 2010).
In this method, in order to obtain a target log from m number of wells 10. Application of neural network for sand volume and reservoir
and n number of attributes, the method needs to find the attributes which property estimation
yield the best prediction of the target log after combination in a lowest
least-squared error (Ogiesoba, 2010). The algorithm first searches for a One of the advantages of using the artificial neural network algorithm
single best attribute by exhaustive search of all the attributes, using is related to fast speed and small errors in determining the nonlinear
equation (B-4). The best attribute is the one with the lowest prediction relationships between the input data and the target property (Calderon
error. When the attribute is found, it searches for a second attribute and Castagna, 2007; Hampson et al., 2001; Leiphart and Hart, 2001;
which represents the lowest prediction error when combined with the Pramanik et al., 2004; Walls et al., 2002). Therefore, two types of neural
first best attribute. Next, it tries to find the best triplet of attributes with network including Probabilistic Neural Network (PNN) and Radial Basis
the lowest prediction error, and so on. However, a validation test is Function (RBF) were emphasized to obtain the Vsand and reservoir
needed as a criterion to stop adding additional attributes when an property. However, PNN represented better results in the present study.
optimal number of attributes is found. The Probabilistic neural network is a feedforward neural network,
The validation test involves several steps. The validation error is including Input, Hidden and Output layer. The network is regarded as
computed by hiding a particular well and using the remaining wells and one of the most effective neural networks used to predict the petro-
attributes as the best single attribute during analysis for m number of physical characteristics of seismic attributes by different researchers
wells including the target log and n attributes. This process of cross (Hampson et al., 2001; Li, 2014; Masters, 1995; Pramanik et al., 2004;
validation continues until all the wells can arrive at a total validation Specht, 1990). The data used by the PNN consist of a series of training
error. The process is repeated for the best two attributes, three attributes, samples such as the sample of {A1i, A2i, A3i, Li}, i ¼ 1, 2, 3, … n. For a
and so on. Then, based on a plot of validation error versus the number of given training sample, the PNN assumes that each new output log value
attributes, suitable attributes are identified for prediction. can be expressed as a linear combination of the log values in the training.
Fig. 9 displays the application of multi-attribute regression method In other words, for a new sample, by having a property vector x, the new
for sand volume, effective porosity and water saturation prediction; the log value is estimated as follows:
red line is cross correlation between the actual and the predicted po- Pn
rosities. The numerical errors for each of target logs are shown in Table 1, i¼1 Li exp½ DðX; Xi Þ
LðXÞ ¼ P n (1)
in which each line indicates a different multi-attribute transform, i¼1 exp½ DðX; Xi Þ
including above all its attributes. Based on the results in Fig. 10, for
267
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 13. Sand volume maps of different vertical slices of Asmari reservoir zones. Location of wells used in the field is marked with small black circles. a) 2 ms slice of Asmari Horizon, b)
2 ms slice of Asmari reservoir zone two (Z_2), c) 2 ms slice of Asmari reservoir zone three (Z_3), d) 2 ms slice of Asmari reservoir zone four (Z_4).
where n indicates the number of measurements and D(x, xi) is defined by PNN, which uses a Gaussian kernel estimator instead of the sigmoid
function based on the distance between points in seismic attribute space.
3
X 2
Xj Xij Learning is much faster than other types of networks in RBF, when the
DðX; Xi Þ ¼ (2)
j¼1
σj number of input variables are not very high. However, an increase in the
number of input variables leads to an geometrical increase in the number
The quantity D(x, xi) represents the distance between the point we are of hidden units. Accordingly, the use of this network becomes practically
trying to estimate, x, and the training points, xi. This distance is scaled by impossible for a large number of input variables.
the quantity σ j (smoothing parameter) which may be different for each In practice, using neural networks in this study is divided into
of the attributes. The network training aims to determine σj, and mini- three steps:
mizing the validation error is regarded as the criterion used for deter-
mining σj. Thus, the validation result for the mth target sample is 1) The selection of appropriate attributes by applying the stepwise
as follows: regression and its validation
P 2) Neural network training to derive the statistical relationship between
Li exp½ DðXm ; Xi Þ
b
L m ðXm Þ ¼ Pi≠m
n (3) seismic attributes and target logs at well locations
i≠m exp½ DðXm ; Xi Þ 3) The application of derived neural networks to the 3D seismic data
volume
Radial basis function artificial neural networks (RBF) are similar to
268
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 14. Arbitrary line from porosity cube created using multi attribute regression method. Porosity decreases from top to bottom of reservoir. Porosity logs are inserted at the well
locations to compare the actual porosity with the predicted porosity near the well locations. Map of the arbitrary line is shown in the upper right of the figure.
The two aforementioned neural network was trained by using the shows medium acoustic impedance and effective porosity values, and is
attributes obtained from multi-attribute regression for Vsand, water mostly composed of fine to very fine-grained sandstones with dolomitic
saturation and effective porosity. After training, PNN and RBF provided a cement and argillaceous matrix and dolomitic skeletal grainstone with
correlation coefficient of 0.87 and 0.84, respectively, between actual and biomoldic, vuggy and intercrystalline porosity (Fig. 11b and c). Cluster 3
predicted sand volume, which is better than multilinear regression compared with the aforementioned clusters characterized by high
(Table 2). However, the correlation coefficient for the validation result acoustic impedance and low average effective porosity. The facies of this
drops to 0.72 and 0.68, which is marginally worse than the validation cluster consist of claystones, mudstones and wackestones with occasional
result for multilinear regression (0.73). dolomitization (Fig. 11d). These facies are deposited in sub-tidal
As it was already mentioned, the result of validation is regarded as the lagoonal and lower shoreface to offshore settings, indicating the lowest
real measure of performance. Therefore, based on this criterion, multiple values of core porosity and permeability.
attribute regression method used for Vsand prediction is more accurate
than other methods in this case. 12. Results and discussion
Similarly, based on the cross-correlation between actual and pre-
dicted water saturation in training set and validation tests, the prediction Based on both linear and non-linear algorithms on 3D seismic attri-
power increases through using neural network, compared with a single butes and log data of the Mansuri field, the multi-attribute analysis as a
attribute and the multi-attribute regression analysis. In other words, the suitable method was used for predicting sand volume and effective
neural network can sufficiently improve the performance of the water porosity. Further, neural network was selected to estimate water
saturation estimation in Mansuri oil field. In addition, based on the saturation.
validation tests, multi-attribute regression was used as an efficient and Based study of core and thin sections of Asmari reservoir, the
practical method for effective porosity prediction of Asmari reservoir in impedance data can be used to determine the lateral and vertical varia-
the Mansuri field. tions in lithology and facies units. The inversion result indicated that low-
amplitude reflections in the upper part of the reservoir are due to low-
11. Geological reservoir characterization based on acoustic impedance (soft) layers, as they are mostly considered as sandstones li-
impedance thology. The correlation of low impedance with sandstone intervals can
be explained by the lower velocity of unconsolidated channel sands,
In a complementary study, k-means clustering was utilized to char- compared to dolomitic skeletal grainstone, packstone and Lime/dolo-
acterize and classify reservoir based on acoustic impedance. Accordingly, mitic mudstone/wackestone in this field. Furthermore, high acoustic
Asmari reservoir was classified into three main clusters including low (AI impedance value (Purple color) is congruent with low porous and
<7926), medium (7930 < AI < 10610) and high (AI >10610) acoustic consolidated Lime/dolomitic mudstone/wackestone facies in the studied
impedance. Table 3 indicates the main characteristics of these three reservoir (Fig. 8). Accordingly, seismic inversion for acoustic impedance
clusters or groups. Then, regarding the core and thin sections of reservoir, is a useful tool in reservoir characterization of Asmari reservoir in the
the difference in AI value can be described based on petrophysical and field in order to yield lithology and reservoir property.
geological characteristics of the reservoir in the field under study. A multi-attribute transform method was applied to the 3-D seismic
Accordingly, cluster1 encompasses more unconsolidated fine to coarse- volumes to generate sand probability volumes. Figs. 12 and 13 illustrate
grained channel sands and occasionally consolidated medium to arbitrary line and different slices with 2 ms thickness from the top of each
coarse-grained sandstones cemented with dolomite (Fig. 11a). Cluster 2 zone in the studied field, respectively. The results demonstrated that
269
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Fig. 15. Effective porosity maps of different vertical slices of Asmari reservoir zones. Location of wells used in the field is marked with small black circles. a) 2 ms slice of Asmari Horizon,
b) 2 ms slice of Asmari reservoir zone two (Z_2), c) 2 ms slice of Asmari reservoir zone three (Z_3), d) 2 ms slice of Asmari reservoir zone four (Z_4). Note concentration of high porosity
values in central and east of survey area.
siliciclastics are best developed in the west, and to the bottom of the and are concentrated mostly in the along central and east part of
reservoir, and it gradually increase towards the east and the north of the the reservoir.
field. Geometrically, sandstones are more deposited as the thickness of Based on the results of multi-regression analyses, in order to predict
accumulation of continuous sands, which is locally incisions, reac- water saturation, the first four attributes of time, acoustic impedance,
tivations possible. filter 5/10–15/20 and cosine Instantaneous Phase, could be considered
Based on the results of this study, a multi-attribute regression can as the optimal inputs for predicting water saturation using a neural
successfully predict effective porosity from the use of acoustic imped- network (PNN) (Table 1). Fig. 16 indicates a cross plot of predicted water
ance, Filter 5/10–15/20, Amplitude Envelope, Time, Integrated Absolute saturation values against actual water saturation values using points of all
Amplitude and Cosine Instantaneous Phase attributes. seven wells. It was evident that PNN technique provided a better corre-
Based on representative arbitrary line and horizon slices through the lation coefficient (0.91) and average error (0.02126), compared with
effective porosity volume (Figs. 14 and 15), high-porosity zones is more multi-attribute linear regression. As the neural network can better
corresponding to high sand volume parts. Four different horizon slices determine the nonlinear relationship between seismic attributes and
(Fig. 15) below the top of the Asmari show that overall effective porosity water saturation log, the neural network result was applied to the entire
decreases with depth, as it was found from inversion results in Fig. 7. In survey, starting from the top of the Asmari to the base of the Pabdeh in
addition, the most porous zones are more continuous at shallow levels order to generate water saturation volume. The water saturation maps
270
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
13. Conclusion
Fig. 17. Water saturation maps of different vertical slices of Asmari reservoir production zones. Location of wells used in the field is marked with small black circles. a) 2 ms slice of Asmari
Horizon, b) 2 ms slice of Asmari reservoir zone two (Z_2), c) 2 ms slice of Asmari reservoir zone three (Z_3), d) 2 ms slice of Asmari reservoir zone four (Z_4).
271
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
regarded as a key parameter for directing geologic interpretation and field. Accordingly, the best porous and permeable reservoir intervals
detecting lateral lithologic variation. Based on the results, the upper parts with low AI were found in the unconsolidated fine to medium-grained
of Asmari reservoir have lower AI values than the lower reservoir because channel sandstones, while high AI value was compatible with low
it is more sandy and porous. Then, a linear regression and neural network porous and mud-supported facies and claystone. Hence, for this reservoir
methods were conducted in a volume-based approach to model Vsand and case, impedance data may be used to predict depositional facies away
reservoir property distribution by using seismic attributes in addition to from the already drilled wells.
the inversion attribute and well log. The present study provided a better understanding of heterogeneity
The results indicated the prediction of effective porosity and sand and quality of reservoir, which plays a pivotal role in analyzing the
volume obtained from multi-attribute regression analysis. Also, the various producing units in the study area, which is helpful in developing
neural network is the best method for estimating water saturation. drilling for optimal hydrocarbon recovery.
Representative transects and horizon slices through the effective
porosity, sand volume and water saturation volumes show that high Acknowledgements
effective porosity zones are more related to high sand volume parts.
Further these parameters best developed in the west, gradually increased This contribution represents part of the Ph.D. thesis of first author at
towards the east and the north of the field. Moreover, water saturation in Shahid Chamran University of Ahvaz, Iran. The authors acknowledge the
the central and east parts of the reservoir is low, which corresponds to the National Iranian South Oil Company (NISOC) for providing the subsur-
hydrocarbon-bearing area. face data and for permission to publish the data. The authors would also
Regarding the correlation between core data and AI, variations in like to thank the editor and the anonymous reviewers for the critical and
acoustic impedance were more related to geological (depositional, constructive comments which greatly contributed to the improvement of
diagenetic) and petrophysical characteristics of Asmari reservoir in the the manuscript.
Convolutional model is regarded as the basic premises behind all seismic inversion methods, upon which the seismic trace(t) is simply the
convolution(*) of the earth's reflectivity with a seismic source function w(t) by adding a noise component n(t) (Russell and Hampson, 1991):
where ri represents the reflection coefficient at the ith geological interface, and zi indicates the impedance at the ith geological layer and is defined as the
product of compressional velocity (v) and density (p). By inverting equation (A-2), the impedance value of each layer can be recursively derived from
the impedance value of the above layer. The recursive inversion equation is written as follows:
1 þ ri
ziþ1 ¼ Zi (A-3)
1 ri
Equation (A-3) is generalized to provide the impedance value of an arbitrary M layer by
M
Y
1 þ ri
ZM ¼ Z1 (A-4)
i¼2
1 ri
Based on the logarithm of the above equation, a linear approximation was used:
XM
r3 r5
lnðZM Þ ¼ lnðZ1 Þ þ 2 ri þ i þ i þ … ; (A-5)
i¼2
3 5
As it is possible to discard the high-order terms in the summation, equation(A-5) is rewritten as follows:
!
X
M
ZM ¼ Z1 exp 2 ri ; (A-6)
i¼2
The preceding approximation is used in recursive inversion for the transformation of reflectivity into impedance and is valid if ri <(0.3) (Leite and
Vidal, 2011; Oldenburg et al., 1983; Russell, 2004).
In order to use the recursive inversion properly and obtain absolute impedance, the noise component should be removed or reduced and it should
deconvolve the seismic traces into the reflectivity series. To accomplish this, a Model-based inversion is applied for minimizing the objective function
(Russell, 2004).
272
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
In most cases, the function to model the reservoir parameters via m different attributes is written as follows:
T ¼ w0 þ w1 A1 þ w2 A2 þ … þ wm Am (B-1)
where T represents reservoir parameters (target parameter), Ai is the ith seismic attribute and wi shows the weighting factor.
The least squares method is used to get the target parameter:
P P P X
2 3 2 N 31 2 L 3
w0 X P A1i P A2i P Ami X i
6 w1 7 6 X A1i A2 A1i A2i A1i Ami 7 6 A L 7
6 7¼6 P 1i P 2 P 7 6 X 1i i 7 (B-2)
4 w2 5 4 A A1i A2i A A2i Ami 5 4 A2i L2i 5
X 2i P P 2i P 2 X
wm Ami A1i Ami A2i Ami Ami Ami L2i
It is worth noting that one-to-one corresponding samples may not be optimal due to the difference in the frequency between target log and seismic
attributes. A convolution operator is used to solve the problem. Based on this operator, each sample of the log is related to a group of neighboring
samples on the seismic attribute (Hampson et al., 2001).
The extension of equation (B-1) to involve the convolutional operator is as follows:
T ¼ w0 þ w1 * A1 þ w2 * A2 þ … þ wm * Am (B-3)
where * represents the convolution operator and wi indicates the operators of a specified length.
Similarly, the convolution operator is computed by the least squares method:
1 XN 2
E2 ¼ Li w0 þ w1 * A1i þ w2 * A2i þ … þ wm * Ami (B-4)
N i¼1
As given in equation (B-4), the optimal operator determined by minimizing the mean-squared error between the actual and the predicted values of
the target log.
References Lavergne, M., Willm, C., 1977. Inversion of seismograms and pseudo velocity logs.
Geophys. Prospect 25, 231–250.
Leiphart, D.J., Hart, B.S., 2001. Comparison of linear regression and a probabilistic neural
Adekanle, A., Enikanselu, P.A., 2013. Porosity prediction from seismic inversion
network to predict porosity from 3-D seismic attributes in Lower Brushy Canyon
properties over “XLD”Field. Niger. Delta. Am. J. Sci. Ind. Res. 4, 31–35.
channeled sandstones, southeast New Mexico. Geophysics 66, 1349–1358.
Bosch, M., Mukerji, T., Gonzalez, E.F., 2010. Seismic inversion for reservoir properties
Leite, E.P., Vidal, A.C., 2011. 3D porosity prediction from seismic inversion and neural
combining statistical rock physics and geostatistics: a review. Geophysics 75,
networks. Comput. Geosci. 37, 1174–1180.
75A165–75A176.
Li, M., 2014. Geophysical Exploration Technology: Applications in Lithological and
Calderon, J.E., Castagna, J., 2007. Porosity and lithologic estimation using rock physics
Stratigraphic Reservoirs. Elsevier.
and multi-attribute transforms in Balcon Field, Colombia. Lead. Edge 26, 142–150.
Lindseth, R.O., 1976. Seislog process uses seismic reflection traces. Oil Gas. J. 74, 67–71.
Draper, N.R., Smith, H., Pownell, E., 1966. Applied Regression Analysis. Wiley, New York.
Masters, T., 1995. Advanced Algorithms for Neural Networks: a Cþþ Sourcebook. John
EDCO Kish International Company, 2006. Geological Studies Report for Mansuri Oil Field
Wiley & Sons, Inc.
Development in Asmari Formation.
Na’imi, S.R., Shadizadeh, S.R., Riahi, M.A., Mirzakhanian, M., 2014. Estimation of
Farfour, M., Yoon, W.J., Kim, J., 2015. Seismic attributes and acoustic impedance
reservoir porosity and water saturation based on seismic attributes using support
inversion in interpretation of complex hydrocarbon reservoirs. J. Appl. Geophys. 114,
vector regression approach. J. Appl. Geophys. 107, 93–101.
68–80.
Naeem, M., El-Araby, H.M., Khalil, M.K., Jafri, M.K., Khan, F., 2015. Integrated study of
Hampson, D.P., Schuelke, J.S., Quirein, J.A., 2001. Use of multiattribute transforms to
seismic and well data for porosity estimation using multi-attribute transforms: a case
predict log properties from seismic data. Geophysics 66, 220–236.
study of Boonsville Field, Fort Worth Basin, Texas, USA. Arab. J. Geosci. 8,
Hart, B.S., Balch, R.S., 2000. Approaches to defining reservoir physical properties from 3-
8777–8793.
D seismic attributes with limited well control: an example from the Jurassic
Ogiesoba, O.C., 2010. Porosity prediction from seismic attributes of the Ordovician
Smackover Formation, Alabama. Geophysics 65, 368–376.
Trenton-Black River groups, Rochester field, southern Ontario. Am. Assoc. Pet. Geol.
Huuse, M., Feary, D.A., 2005. Seismic inversion for acoustic impedance and porosity of
Bull. 94, 1673–1693.
Cenozoic cool-water carbonates on the upper continental slope of the Great
Oldenburg, D.W., Scheuer, T., Levy, S., 1983. Recovery of the acoustic impedance from
Australian Bight. Mar. Geol. 215, 123–134.
reflection seismograms. Geophysics 48, 1318–1337.
Iturraran-Viveros, U., Parra, J.O., 2014. Artificial Neural Networks applied to estimate
Pearson, R., Hart, B., 1999. Convergence of 3-D seismic attribute-based reservoir property
permeability, porosity and intrinsic attenuation using seismic attributes and well-log
prediction and geologic interpretation as a risk reduction tool: a case study from a
data. J. Appl. Geophys. 107, 45–54.
Permian intraslope basin. In: Internat. Mtg., Soc. Expl. Geophys, pp. 896–899.
James, G.A., Wynd, J.G., 1965. Stratigraphic nomenclature of Iranian Oil Consortium
Expanded Abstracts.
Agreement Area. AAPG Bull. 49 (12), 2182–2245.
Perez-Mu~ noz, T., Velasco-Hernandez, J., Hernandez-Martinez, E., 2013. Wavelet
Kadkhodaie-Ilkhchi, A., Rezaee, M.R., Rahimpour-Bonab, H., Chehrazi, A., 2009.
transform analysis for lithological characteristics identification in siliciclastic oil
Petrophysical data prediction from seismic attributes using committee fuzzy
fields. J. Appl. Geophys. 98, 298–308.
inference system. Comput. Geosci. 35, 2314–2330.
Pramanik, A.G., Singh, V., Vig, R., Srivastava, A.K., Tiwary, D.N., 2004. Estimation of
Kadkhodaie-Ilkhchi, R., Moussavi-Harami, R., Rezaee, R., Nabi-Bidhendi, M.,
effective porosity using geostatistics and multiattribute transforms: a case study.
Kadkhodaie-Ilkhchi, A., 2014. Seismic inversion and attributes analysis for porosity
Geophysics 69, 352–372.
evaluation of the tight gas sandstones of the Whicher Range field in the Perth Basin,
Raeesi, M., Moradzadeh, A., Ardejani, F.D., Rahimi, M., 2012. Classification and
Western Australia. J. Nat. Gas. Sci. Eng. 21, 1073–1083.
identification of hydrocarbon reservoir lithofacies and their heterogeneity using
Kadkhodaie-Ilkhchi, R., Rezaee, R., Moussavi-Harami, R., Kadkhodaie-Ilkhchi, A., 2013.
seismic attributes, logs data and artificial neural networks. J. Pet. Sci. Eng. 82,
Analysis of the reservoir electrofacies in the framework of hydraulic flow units in the
151–165.
whicher range field, perth basin, western Australia. J. Pet. Sci. Eng. 111, 106–120.
Research Institute of Petroleum Industry (RIPI), 2006. Unpublished Sedimentology and
Khoshdel, H., Riahi, M.A., 2011. Multi attribute transform and neural network in porosity
Geological Reservoir Characterization of the Asmari Formation in the Mansouri Field.
estimation of an offshore oil field—a case study. J. Pet. Sci. Eng. 78, 740–747.
Russell, B., Hampson, D., 1991. Comparison of poststack seismic inversion methods. In:
Kumar, R., Das, B., Chatterjee, R., Sain, K., 2016. A methodology of porosity estimation
1991 SEG Annual Meeting. Society of Exploration Geophysicists.
from inversion of post-stack seismic data. J. Nat. Gas. Sci. Eng. 28, 356–364.
Russell, B.H., 2004. The Application of Multivariate Statistics and Neural Networks to the
Latimer, R.B., Davidson, R., Van Riel, P., 2000. An interpreter's guide to understanding
Prediction of Reservoir Parameters Using Seismic Attributes. University of Calgary,
and working with seismic-derived acoustic impedance data. Lead. Edge 19, 242–256.
Alberta.
273
I. Zahmatkesh et al. Journal of Petroleum Science and Engineering 161 (2018) 259–274
Russell, B.H., 1988. Introduction to Seismic Inversion Methods. Society of Exploration Stocklin, J., 1968. Structural history and tectonics of Iran: a review. Am. Assoc. Pet. Geol.
Geophysicists. Bull. 52, 1229–1258.
Saltzer, R., Finn, C., Burtz, O., 2005. Predicting V shale and porosity using cascaded Walls, J.D., Taner, M.T., Taylor, G., Smith, M., Carr, M., Derzhi, N., Drummond, J.,
seismic and rock physics inversion. Lead. Edge 24, 732–736. McGuire, D., Morris, S., Bregar, J., 2002. Seismic reservoir characterization of a US
Simm, R., Bacon, M., Bacon, M., 2014. Seismic Amplitude: an Interpreter's Handbook. Midcontinent fluvial system using rock physics, poststack seismic attributes, and
Cambridge University Press. neural networks. Lead. Edge 21, 428–436.
Snedden, J.W., 2013. Channel-body basal scours: observations from 3D seismic and Yao, T., Journel, A.G., 2000. Integrating seismic attribute maps and well logs for porosity
importance for subsurface reservoir connectivity. Mar. Pet. Geol. 39, 150–163. modeling in a west Texas carbonate reservoir: addressing the scale and precision
Soubotcheva, N., Stewart, N., 2006. Predicting lithology and porosity at the pikes peak problem. J. Pet. Sci. Eng. 28, 65–79.
heavy oilfield, Saskatchewan using 3D seismic data and well logs. Can. Soc. Explor. Zonneveld, J.-P., Gingras, M.K., Pemberton, S.G., 2001. Trace fossil assemblages in a
Geophys. Rec. 35–39. Middle Triassic mixed siliciclastic-carbonate marginal marine depositional system,
Specht, D.F., 1990. Probabilistic neural networks. Neural Netw. 3, 109–118. British Columbia. Palaeogeogr. Palaeoclimatol. Palaeoecol. 166, 249–276.
274