You are on page 1of 8

Estimation of Reservoir Quality by Attribute Integration Through Fuzzy Logic

Muhammad M. Saggaf, Saudi Aramco, Saudi Arabia Abstract


We describe an approach for estimating the reservoir potential by integrating multiple attributes derived from various input sources. We present both a simple heuristic method and one based on fuzzy logic, and show the two are effectively identical. Both methods can be implemented in the forward sense (without utilizing knowledge from existing wells) or the inverse sense (using existing well information to condition the final outcome). We applied the method to assess the reservoir quality throughout the Haradh area of the Ghawar Field of Saudi Arabia and evaluated the validity of the study by analyzing the outcome of wells drilled on the basis of this analysis.

Introduction
For optimal placement of a development well, a decision-maker is faced with a multitude of attributes geophysical, geological, or engineering that have varying degrees of confidence and that are related to the drilling objective in varying extents. These attributes are oftentimes large in number; thus, qualitatively assessing their relative impacts on the development problem can be laborious, subjective, or worse, erroneous. Sometimes they are incomplete or even contradictory in their inference to the expected reservoir quality. None of them, singly and separately, predicts the occurrence or amount of hydrocarbon-bearing reservoir with such an overwhelming confidence that precludes the need for the others. In reality, each attribute may sense the same or a different reservoir property than another. Therefore, it is the combination or integration of these attributes that yields a more reliable indicator of reservoir quality than any single attribute can. The goal then is to integrate a multitude of fragmented pieces of knowledge derived from various input sources to come up with a single number at each location in the field that is indicative of the reservoir potential at that particular location. In addition to furnishing an accurate means of assessing the reservoir potential at any point in the field, this approach also provides a permanent record of why a particular location was selected over another, what pieces of information (attributes) entered into the making of that decision, and to what extent each of these attributes influenced that decision. In effect, then, this approach renders the entire decision-making process more objective and more open, and makes it quite facile and convenient to revisit that process after the well has been drilled to assess the decision and re-evaluate the contribution of each of the attributes into that decision and perhaps readjust those contributions to fine-tune the method to yield a more accurate estimate of the reservoir potential for further wells to be drilled. In this paper, we first describe a heuristic method of integrating these multiple attributes to estimate the reservoir potential. Then, we extend this method to utilize information from existing drilled wells to invert for the optimum combination of these attributes based on the a priori information obtained at the wells. Finally, we describe a fuzzy-logic model to achieve the same stated objective and show that it is effectively equivalent to the heuristic method in terms of its implementation and results. This lends the heuristic method a firm theoretical foundation and demonstrates its analytical robustness. We then conclude by commenting on the quality of the wells that were drilled on the basis of the reservoir potential estimated by the method described in this paper. Before discussing the method, we begin by describing the field on which this method was applied, the existing wells in that field, and the various attributes utilized in this study.

Field Description
We apply the method to an actual case over the Haradh area of the Ghawar Field of Saudi Arabia. The reservoir target in the development drilling program is the gas-bearing Unayzah sandstone reservoir, a Lower Permian shallow-marine delta that progrades to the east. The top of the reservoir is

at about 12000 feet below sea level. This reservoir consists of a coarsening-upward sequence of sands with interlayers of silts and shale that has a gross thickness of between 200-300 feet. Well data indicates that the best-quality reservoirs are confined to the uppermost layers of the sequence and that the reservoir quality (net-to-gross ratio) improves towards the eastern flank of the field. A 3D seismic survey has been acquired over this field and the base of the overlying Khuff carbonates produces a strong, mappable reflection throughout the study area. From this seismic reference, the top of the Unayzah reservoir is obtained by isopaching down by 60-75 feet. Based on five wells as control, a window bracketing the reservoir was inverted using two inversion schemes a model-based scheme that uses a single (representative) wavelet and another method that accepts a laterally varying wavelet. Of the five wells drilled at the time of the study, one (Well A) penetrated the delta plain and encountered very thin sands containing little or no gas. About 10 km to the east, a second well (Well B) encountered a 320-foot interval, three quarters of which is gas-filled sandstone. Five kilometers farther east, at Well C, an equally thick sand reservoir interval was penetrated, but which was completely water bearing. Two other wells, D and E are north and south, respectively of Well B. Both of these wells encountered gas-bearing reservoirs of less quality (lower net-to-gross ratios) than either Wells B or C. Thus, these wells sample the spectrum of the reservoir, but they are widely separated from each other. Moreover, they do not adequately describe the short-range variability of the reservoir, particularly as the reservoir deteriorates in quality towards the west. Furthermore, the exact gas-water contact has not been confirmed by any of the wells, although it must lie between the lowest-known gas and the highest-known water. The integration of multiple attributes aims to maximize the likelihood of encountering sufficiently thick sand reservoir containing gas and provide a way to rank a series of development well candidates for the drilling program.

Input Attributes
By attributes, we mean here any map expressing a particular property of the reservoir, derived from any input source seismic, geologic, engineering, or otherwise. In other words, we use the general sense of the word here rather than the more restricted sense that pertains only to complex trace attributes derived from seismic data. Maps of several attributes related to the reservoir and its contained hydrocarbon exist in this field from various sources and disciplines. The field geologist has made a contour of the net-sand based on well data and his concept of the deposition of the delta. He has also provided average values of the porosity-thickness and gas volume per unit area at each well. Average impedance maps of the reservoir interval have been calculated from the two separate inversion runs mentioned above. A map of the RMS amplitude mismatch between actual seismic trace and inversion synthetic trace has also been computed to indicate where inversion has converged on a solution and where it has not quite succeeded. Seismic attribute analysis has also yielded a map of the average instantaneous frequency within the reservoir interval. Seismic traces windowed over the reservoir have been extracted from the 3D cube and classified into four facies classes using unsupervised vector quantization by competitive neural networks (Saggaf, 20001). Some of these attributes, such as net sand, relate directly to the reservoir, but it assumes continuity between wells. Another attribute, inversion error, does not relate to the reservoir but it qualifies the reliability of a separate attribute, the acoustic impedance from inversion, which is related to the average porosity of the reservoir. Still another attribute, seismic facies, simply compares the similarity of the seismic response at any point on the field with that representative of the facies observed at any existing well location, assuming that similar seismic traces likely, but not necessarily, reflect the same geologic layering. Thus, each attribute contributes to the understanding of the reservoir or evaluation of the data about the reservoir. From the decision-makers point of view, one would expectedly wish to locate a development well where: 1. Average impedance (from both inversion methods) is low since porous sands have relatively low velocities and densities, 2. Average instantaneous frequency is low due to absorption by the gas,

3. Net sand is interpreted to be thick, 4. Seismic facies is similar to the productive Well B and dissimilar to the unproductive Well A, 5. Error in the inversion is small. With these statements, a decision system is tacitly expressed, which is an unquantified amalgamation of the preferences of the decision-maker. One may further modify this system by varying the weights mentally assigned to each attribute. For example, we may rely heavily on the average impedance as it is a direct measurement of the reservoirs response to the transmitted seismic energy, but, perhaps, less on average frequency as this can be affected by non-reservoir conditions. Depending on whether we agree with the geologist as to the continuity of these sands in a deltaic setting, we may put heavy emphasis on the net sand map for distal locations as compared to more proximal locations, i.e., closer to the delta plain. In any case, we quickly find that there is a significant difficulty sorting through the preferences and relative emphasis placed on each attribute in this qualitative analysis as the number of candidate wells increases.

Heuristic Method
Heuristically, we systematize and quantify the decision style described above by considering how each attribute separately relates to the favorability of the reservoir. Then, we assess how we view the relative importance of these various indicators by assigning weights to them. Finally, we combine the weighted attributes in a fashion that reflects our decision preference. The overall reservoir potential is thus given by:

P ( x, y ) = wi Ri ( x, y ) ,
i =1

(1)

where Ri represents each of the individual reservoir indicators derived from the raw attributes and wi is the weight assigned to each of these indicators. Thus, in order to compute the overall reservoir potential, we need to compute the individual reservoir indicators, Ri, from each input attribute and decide on the weight we assign to each of these indicators. The weights are assumed to be constant for each indicator throughout the field and do not vary laterally, although this restriction can be relaxed, if desired.

1 0.9 0.8 0.7 0.6

1 0.9 0.8 0.7 0.6

1 0.9 0.8 0.7 0.6

1 0.9 0.8 0.7 0.6

1 0.9 0.8 0.7 0.6

1 0.9 0.8 0.7 0.6

RQ

RQ

RQ

RQ

RQ

0.5 0.4 0.3 0.2 0.1 0 1.5 2 2.5 3 3.5 4 4.5 5 5.5 x 10
4

0.5 0.4 0.3 0.2 0.1 0 1.5 2 2.5 3 3.5 4 4.5 5 5.5 x 10
4

0.5 0.4 0.3 0.2 0.1 0 0 10 20 30 40 50 60

0.5 0.4 0.3 0.2 0.1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

0.5 0.4 0.3 0.2 0.1 0 0 50 100 150 200 250

RQ

0.5 0.4 0.3 0.2 0.1 0 0 1000 2000 3000 4000 5000 6000 7000 8000

Impedance

Impedance

Avg. Frequency (Hz)

Seismic Facies

Net Sand Thickness (ft)

Inversion Error

min = 20,000 max = 51,500 RI = -0.031*x / 1000 + 1.625

min = 20,000 max = 51,500 RI = -0.031*x / 1000 + 1.625

min = 10 max = 55 RI = -0.018*x + 1.18

min = 1000 max = 9000 RI = 0.33*x - 0.33

min = 0 max = 235 RI = 0.004*x

min = 1000 max = 9000 RI = 0.143*x / 1000

Figure 1: Reservoir indicator (left) and transform used to obtain it (middle and right), for (from top): impedance 1, impedance 2, average frequency, seismic facies, net sand thickness, and inversion error. The individual reservoir indicators are computed through transforms applied to the original attributes to convert them into measures of the degree of favorability of the reservoir at each location in the field. These indicators thus range from 0 (poor reservoir) to 1 (excellent reservoir potential). For example, when acoustic impedance is low, its corresponding indicator would be high (close to 1), since low impedance is indicative of better porosity development (better reservoir potential), and similarly for

each of the remaining attributes. The transforms are constructed in a simple manner. For each attribute, its minimum and maximum are computed, then a linear map (with either a positive or a negative slope, depending on the particular attribute) is formed such that it maps that attribute into the interval [0,1], giving rise to its corresponding reservoir indicator. Figure 1 shows a map of each attribute and the transform utilized to convert it to an appropriate reservoir indicator. It should be noted that other, non-linear, transforms can also be utilized. We comment on that later in the paper. Determining the weights can be performed in either the forward or inverse sense. In the forward sense, the weights are selected subjectively according to the overall decision preference and style of the decision-maker. The words style and preference are used here to indicate that different decision-makers view the attributes and their significance differently with respect to the same decision objective that of drilling a development well. Res. Indicator Forward Weight Inverted Weight However, by quantifying the preferences, consensus among decision participants may be Impedance 1 1.0 0.92 reached in stages, attribute by attribute, weight by Impedance 2 1.0 1.37 weight. Final consensus on the overall combination of these preferences constitutes the Frequency 0.5 0.28 agreed-upon decision system. This system may Facies 0.5 0.21 then compute from all the various input attributes a Net Sand 1.0 1.39 single map that reflects the decision-makers or decision groups consensus as to the desirability, Inv. Error 0.7 0.45 favorability, or potential or the reservoir at each location in the field. Table 1 shows the weights Table 1: Original and inverted weights for assigned to each reservoir indicator in the forward each reservoir indicator. method, while Figure 2 shows the final reservoir potential map. Therefore, knowledge from existing wells is not utilized directly in the forward method. In order to use this a priori knowledge from wells already drilled, the inverse approach may be employed. In this approach, a reservoir potential is assigned a priori to each existing well by evaluating the reservoir quality at the existing wells and then optimal weights are calculated such that the resulting computed overall reservoir potential matches the values assigned a priori at the wells as closely as possible. Assume we have M existing wells, let Pj be the a priori reservoir potential assigned to existing well j, and let Rji be the ith reservoir indicator evaluated at well j, then from (1) we have:

Pj = wi R ji ,
i =1

j = 1K M
where:

(2)

or

Rw = P, R11 R R = 21 M R M1 R12 R22 M RM 2 R1N P1 L R2 N P2 , P = M M P L RMN M L w1 w2 , and w = M w N

(3)

(4)

P is known from the a priori values and R is computed at each well by transforming the original attributes into reservoir indicators. Thus, the unknown weights w can be readily calculated by solving the system of equations (3). The system is solved either directly if the number of existing wells is less than or equal to the number weights (number of attributes), where in the former case it will be non-unique as the system would be under-determined, or by least-squares fitting if the number of wells is greater than the number of weights (over-determined system). In practice, however, the system is almost always over-determined, since the nine closest traces to each well (one trace at the center and eight at the perimeter of a square containing the well) are utilized instead of just a singe trace closest to the well, to regularize the system and make the solution more robust.

Res. Potential

Figure 3 shows the overall reservoir potential computed at each well (nine values for each well, as described above) from the forward method (note that Well C shows a high reservoir potential even though it is completely wet, since being below the GWC was not part of our criteria for the reservoir potential). For the a priori reservoir potential at each well, we use a single value for all nine traces (computed as the average of the nine values obtained for that well from the forward method). These values are shown on Figure 3. We then invert for the optimal weights in the manner described above. Table 1 shows the resulting weights due to this inverse method. As expected, the inverted weights are close o the original ones utilized in the forward method. Also, using these inverted weights gives rise to reservoir potential values at the wells that match the ones obtained previously (Figure 3 compares the two sets), since the a priori values utilized in the inverse method were largely based on the values derived from the forward method. Thus, the two methods give similar results overall when the same style and preference (consistent a priori information) is utilized in both, and the inverse method can therefore be used to refine and fine-tune the decision-making system to be more objective.

Figure 2: Final reservoir potential map.

0.65 0.6 0.55 0.5 0.45 0.4 0.35 Forward A Priori Inverse

Fuzzy Logic Method


In contrast to binary-valued (bivalent) logic, in which a statement can be either true or false, and hence its truth is ascribed either 0 or 1, multivalued (multivalent) logic can ascribe any number in the interval [0,1] to represent the degree of truth of a statement. This is quite a normal extension of bivalent logic that we humans practice naturally. For example, we often use sentences like the weather is hot or she is young in the relative sense, where the truth of such statements spans an extended gray interval instead of being absolutely true or false (black or white). The second sentence, for instance, is valid when someone is 20 or 30 years old. However, it is unequivocally more valid (more truthful) in the former case. Therefore, we may say that the statement is true to degrees of 0.9 and 0.8 in the two cases, respectively. Although fuzzy logic is sometimes used as a synonym for multivalent logic, its more common use is to describe the logic of fuzzy sets sets that have no crisp, well-defined boundaries, and which may have elements of partial, rather than full, membership (Zadeh, 19652). For fuzzy sets, elements are characterized by a membership function that describes the extent of membership (or the degree of fit) of each element to the set by a number in the interval [0,1]. For example, for the set of young people, the membership function

0.3 0.25 0.2 A>

B>

C>

D>

E>

Wells

Figure 3: For the five wells of the study, the reservoir potential for nine traces near each well (solid), the a priori potential used in the inverse method (dashed), and the potential produced by the inverse method (gray).

a) 1

Membership

0.8 0.6 0.4 0.2 0 0 10 20 30 40 50 60 70 80 90 100

b) 1

Membership

0.8 0.6 0.4 0.2 0 0 10 20 30 40 50 60 70 80 90 100

Age

Figure 4: (a) Membership function for the fuzzy set of young people, (b) membership function for the crisp set of young people.

assigns 0.9 to people of age 20, and 0.8 to those of age 30. Thus, this set does not have a crisp boundary; members can belong to the set to varying degrees of fit. The membership function is shown on Figure 4a. For each age, it assigns a measure of the degree of fit of that age to the definition of the set (being young). This measure is called the possibility. Note that crisp sets are special cases of fuzzy sets, where the membership function has only two values, 0 and 1. For crisp sets, something can either belong totally to the set or not belong at all; there is no concept of partial membership. To see how fuzzy sets represent a more natural way that mimics how we think, consider what would happen had we used a crisp, rather than a fuzzy, set in our description. We could have defined our set of young people as those of age 30 or less. The equivalent membership function for such a crisp set is shown on Figure 4b, where it simply assigns 1 to people of age 30 or less, and 0 otherwise. However, this definition means that persons of ages 30 and 10 are equally young, and that someone 30 years old is young, but another a day older is not young at all! Such a notion is, of course, absurd. In order to determine the reservoir potential using fuzzy logic, we first describe our decision-making system using simple rules of the form: if net sand is thick and frequency is attenuated and then the reservoir is good. if net sand is thin and frequency is not attenuated and then the reservoir is poor. etc. Membership functions are assigned to each input predicate (e.g. net sand is thick) of every rule as well as to each possible outcome (e.g. the reservoir is good). To apply this fuzzy decision system, each membership function is evaluated at every location in the field. For each rule, then, this produces an implication into the output membership function. A final decision is made by comparing the values of these functions and defuzzifying the outcome to assess the overall reservoir potential. More details are given by Chen et al. (19953). This system thus entails a separate membership function for each input predicate and each possible outcome, such as: IMF11 = net sand is thick IMF12 = net sand is medium thick IMF13 = net sand is thin IMF21 = high frequency is significantly attenuated IMF22 = high frequency is moderately attenuated IMF23 = high frequency is not attenuated OMF1 = reservoir is excellent OMF2 = reservoir is good OMF3 = reservoir is poor In the remainder of this paper we make several reductions to this system to simplify it for our particular case. 1. We replace expressions like net sand is thick by net sand thickness is high. This allows us to use a single membership function for all input predicates relating to net sand thickness, and similarly for the others. Thus, a single membership function is utilized for net sand, for example, and that function is a measure of the extent of truth of that predicate, or the degree to which that attribute relate to the favorability of the reservoir. For example, when the net sand is thick, the value of the membership function will be high, whereas it would be low when the sand is thin. Such a function would thus look just like the one shown on Figure 1. This significantly cuts down on the number of membership functions that need to be generated, such that we need only one function for each input predicate.

2. We also use a single membership function for the outcome, in a manner similar to the above. Moreover, since we now have a single membership function for the outcome, we need only one rule: if net sand thickness is high and frequency attenuation is high and then the reservoir is good. 3. The degree of contribution of each predicate to the final outcome (the reliability of each predicate) can either be absorbed into the corresponding membership function of that predicate by scaling that membership function or expressed as an explicit weight. 4. The final outcome need not be defuzzified, since we are interested in one (fuzzy) number representing the quality of the reservoir at each location in the field. With these reductions, which apply here since we have a rather simple case, we see that the fuzzy logic method reduces to be the same the heuristic approach we described above. The fuzzy logic method, however, has a robust theoretical background and a firm analytical foundation, although in implementation it is virtually the same as the heuristic approach. This lends confidence to the heuristic approach and establishes its credence. We also generalize the fuzzy logic method to utilize existing a priori well information in a manner very similar to the procedure described above for the inverse heuristic method. The weights are absorbed implicitly into each membership function or can be expressed explicitly. We assume the latter with no loss of generality, as both assumptions lead to the same result. This is because in the former case we can just express each membership function as weighted (normalized or scaled) version of another membership function: Fi = wi Fi . We express the weights explicitly here, though, so we have:

Pj = wi F ji ,
i =1

j = 1K M

(5)

This is similar to the system of equations encountered previously in (2), and is solved in a similar manner to yield the optimal weights for the inverse fuzzy logic method.

Post-Analysis Results
The result of this study was utilized to determine the optimal drilling locations for two new development wells. These two wells were planned and drilled after the conclusion of this work and were placed largely on the basis of the results shown here. The locations of these two wells are shown on Figure 2, and the wells are denoted by F and G. Both wells penetrated good quality thick reservoir (net sand thickness of 230 and 270 feet, respectively) and are candidates to become good gas producers. The outcome of drilling these two wells validates the results of the method presented in this paper.

Conclusions
We described an approach for integrating multiple attributes derived from various input sources to estimate the reservoir potential at any location in the field. We presented a heuristic approach in both the forward sense (that assumes the weights assigned to each individual reservoir indicator) and the inverse sense (that utilizes knowledge from existing wells to objectively invert for the optimal weights). We also presented a fuzzy logic approach and demonstrated that it is equivalent to the heuristic approach in its implementation. Thus, the simple heuristic approach was shown to have a firm analytical foundation from fuzzy logic. We applied the method to compute the reservoir potential in the Haradh area of the Ghawar Field of Saudi Arabia. Results obtained from wells drilled on the basis of the outcome of this study demonstrate the validity of this approach, as both wells penetrated good quality thick reservoir, as predicted from the analysis.

Acknowledgements
We would like to gratefully acknowledge the contributions of the following colleagues to this study: Greg Douglas, Dave Fell, John Fitzmaurice, Greg Gregory, Jung Kim, Frank Martens, Roger Price, Muhittin Senalp, Hafiz Al-Shammery, Sami Al-Shridi, and Chuck Wagner. We would also like to thank Saudi Aramco for supporting this research and for giving us permission for its publication.

References
1. Saggaf, M. M., An integrated seismic and well log analysis for the estimation of reservoir properties, 2000, Ph.D. Thesis: Massachusetts Institute of Technology. 2. Zadeh, L. A., 1965, Fuzzy sets: Information and Control, vol. 8, 338-353. 3. Chen, H. C., Li, L. H., and Fang, J. H., 1995, Evaluation and ranking of prospects by fuzzy multicriteria decision making paradigm, in Bertrand Braunschweig and Ron Day, Eds, Artificial intelligence in the petroleum industry: Symbolic and computational applications: Editions Technip, 315-342.

You might also like