SPE 69625 IMPROVED RESERVOIR PERMEABILITY MODELS FROM FLOW UNITS AND SOFT COMPUTING TECHNIQUES: A CASE STUDY,

SURIA and REFORMA-LIBERTAD FIELDS, COLOMBIA

R. Soto B.*, Ecopetrol-ICP , F. Torres, American Geoexploration, S. Arango, American Geoexploration, G. Cobaleda, Ecopetrol-ICP
* SPE Member
Copyright 2001, Society of Petroleum Engineers, Inc. This paper was prepared for presentation at the SPE Lation American and Caribbean Petroleum Engineering Conference to be held in Buenos Aires, 25-28 March 2001. This paper was selected for presentation by an SPE Program Committee following review of information contained in an abstract submitted by the author(s). Contents of the paper, as presented, have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material, as presented, does not necessarily reflect any position of the Society of Petroleum Engineers, its officers, or members. Papers presented at SPE meetings are subject to publication review by Editorial Committees of the Society of Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper for commercial purposes without the written consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may not be copied. The abstract must contain conspicuous acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract This paper presents a methodology that looks to solve the inverse problem of predicting reservoir properties on uncored intervals/wells, using soft computing techniques (neural networks and fuzzy logic), multivariate statistical analysis and hydraulic flow unit concepts. Our methodology to improve the prediction of permeability in Suria and Reforma-Libertad fields in Colombia is the following: 1.Data quality control. We apply multivariate statistical analysis for quality control of core and log data: 95% confidence ellipses and Q-Q plots are used for that purpose. 2.Rock type identification. We use poral geometry analysis to identify rock types in cored wells. Then, fuzzy logic, core and log variables are used to develop a rock type model to be used in solving the inverse problem, predict the rock type in uncored intervals/wells. 3. Hydraulic flow unit classification. For that purpose, we use the technique based on a modified Kozeny-Carmen equation to calculate the reservoir quality index, RQI=0.0314(K/) , flow zone indicator, FZI=RQI/(/(1-)) and z = /(1-). The basic idea of hydraulic flow unit (HFU) classification is to identify classes that form unit-slope straight lines on a log-log plot of RQI vs. z with similar but not identical FZI value. Each class or hydraulic flow unit has a mean FZI value at the intercept with z = 1, and a maximum and a minimum FZI values. We use log data and the fuzzy logic rock type variable to develop a neural network FZI model to be used in solving the inverse problem, predict FZI

in uncored intervals/wells. The HFU for each uncored interval is determined with its FZI value that falls on a range between maximum and minimum values of FZI. Finally, permeability is calculated knowing its porosity and FZI values. In the literature, the HFU is first determined by Bayesian inference assigning a probability distribution of log values to each HFU and identifying to which population the given set of log readings most likely belong. Then, permeability is estimated from porosity and mean FZI values ignoring the scatter data for each HFU. Permeability estimations from our approach are compared from other conventional techniques to demonstrate that this is a better way to get confident permeability models and to show how to used soft computing techniques to improve reservoir description. Introduction The degree of success of many oil and gas drilling, completion, and production activities depends on the accuracy of the models used in a reservoir description. Permeability is an important parameter in a heterogeneous reservoir characterization. Formation permeability controls the strategies involving well completion and stimulation, and reservoir management. For a low-permeability reservoir, a hydraulic fracture treatment may be needed to optimize the oil and gas recovery. In other zones, a matrix acid treatment might be more economic. The optimal well-spacing and wellproduction rates are dependent of the formation's permeability values. In a high-permeability formation, we may drill fewer wells to drain the reservoir. Therefore, permeability is a key parameter in any reservoir characterization that governs in great extension its handling and development. Permeability is usually measured in laboratory on core samples. However, most drilled wells are not cored. As a result, models are needed to estimate permeability in uncored but logged wells. This is known as the inverse problem.1, 2. Traditionally permeability is estimated from a linear regression model assuming that a linear relationship exists between core porosity and the logarithm of core permeability. Often in this kind of plots for any porosity, permeability can

R. SOTO B., ET AL.

SPE 69625

vary by several orders of magnitude. The method ignores the scatter of data assuming that it is due to measurement errors or second-order fluctuations in reservoir characteristics3. Many investigators4-7 have noted the inadequacy of this classical approach and have demonstrated that permeability is not only dependent on porosity. For example, whereas porosity is generally independent of grain size, permeability is strongly dependent on grain size. Because permeability is controlled by both depositional characteristics, such as grain-size and sorting, and by diagenetic features, a better approach to predict permeability is to take into account the fundamentals of geology and physics of flow at porous media3. This means to attribute interdependency between permeability and geological variations in a reservoir rock and establish relationships between core-derived pore-throat parameters and log-derived macroscopic petrohysical attributes2. Such relationships are the best if rocks of similar fluid conductivity are identified and grouped together as hydraulic flow units (HFU). It has been shown that the variability of permeability is large among HFU and low within them. Models from HFU offer an improved estimation of permeability than traditional regression models because they describe permeability variations in a reservoir more realistically. Additionally, the development in software and hardware in the last years has allowed obtaining non-linear models to represent better the real work. During the last decade, application of neural networks, fuzzy logic and genetic algorithms has increased exponentially to get more confident non-linear reservoir models8-13. These soft computing techniques, unlike regression analysis, do not require specification of a structural relationship between the inputs and outputs but have the ability to infer general rules and typical patterns from specific examples. This paper presents an improved methodology for permeability calculation using HFU and soft computing techniques. The basic concepts of multivariate statistical analysis, neural networks and fuzzy logic are described. Then, a review of the statistical approach permeability model for Suria and Reforma-Libertad fields in Colombia is presented. Next, we show our methodology making first quality control and preprocessing of data. Second, determining the rock types from poral geometry variables and the HFUs from the Flow Unit Indicators, FZIs. Third, solving the inverse problem to predict rock types and FZI in uncored wells using fuzzy logic concepts and neural networks respectively. Finally, estimating permeability from FZI equation. Multivariate Statistical Analysis We may use multivariate statistical analysis to ensure quality control, preprocess data, reduce the number of independent variables or determine the dominant variables, and check the adequacy of the models. Most of the techniques of multivariate statistics are based on the
Quality Control and Preprocessing of Data.

assumption that the data follow a multivariate normal distribution. The importance of the normal distribution rests on the experience that the sampling distributions of many multivariate statistics are approximately normal, regardless of the form of the parent population; because of a central-limit effect.14, 15 We can use those concepts to check if a data sample belongs to the same population or if there is any outlier. We may make histograms to see whether each variable follows a bell-shaped normal density, or Q-Q plots to assess the assumption of normality. Fig. 1 shows an example of a QQ plot for GR variable. In this case the data do not follow the straight reference line. It may mean that data do not belong to the same population. In this case the data come from three populations, rocks 1, 2 and 3. Fig.2 is a Q-Q plot for the same variable, GR, but rock 3. In this plot we can see that the data follow a normal distribution. So, we can use this concept to prevent us from mixing data of different populations (rock types). We also may construct confidence ellipses of 95% for data quality control by pair of variables. If we use those ellipses, it is assumed that the bivariate normal distribution applies. Fig. 3 shows an example of a 95% confidence ellipse and the arrows on the ellipse could be possible suspicious observations (outliers). When dealing with more than one regressor variable, it is useful to scale the data before generating the multiple regression model coefficients. If all inputs (regressors) to the model are scaled to have equal variance and mean, then their corresponding multiple regression coefficients will indicate precisely how important that regressor is to the model. For example, assume that we found for a specific formation that the logarithm of permeability is a function of porosity and lithology (gamma ray log) with more and less impact on permeability respectively. It means that porosity should affect a multiple regression model more than gamma ray. Eq. 1 could represent an invalid regression model: Log Kcore = 10 * GR(API) + 10 * (FRACTION)
.................................... (1)

In this example, the equation is obtained without scaling the input and output variables, and the GR influences the permeability more than porosity because of the scale of gamma ray (API units) and porosity (fraction). Eq. 2 could represent a valid multiple regression model: (Log Kcore)(SCALED) = 0.01 * GR(SCALED) + 10 * (SCALED)
..... (2)

In this case, porosity and gamma ray are in the same scale, and porosity has more impact upon the permeability because of its coefficient. To take into account this principle, we scale both the inputs (regressors) and targets in the range [-1,1] using the following equation: X(SCALED) = 2 * (X X(MIN) )/(X(MAX) X(MIN)) 1
......................... (3)

where X could be any variable. We found that this scaling is

SPE 69625

IMPROVED RESERVOIR PERMEABILITY MODELS FROM FLOW UNITS AND SOFT COMPUTING TECHNIQUES .

critical to accurate modeling in a multivariate environment. Instrumentation developed in recent years allows us to measure dozens of oilfield variables. However, not all the variables are related to each other or more than one variable may measure the same driving principal. The problem to be solved in this case is dimensionality. How many input variables should we use in each case? Finding the major input parameters to predict any variable is not easy. When this happens, there are some ways to simplify the problem and to determine the dominant input variables. One of the ways is to use cross-correlation coefficient values. A cross-correlation coefficient is a measure of the extent and direction of correlation between two variables. If we were to model any target variable, we would use as input variables those that have possible correlation with that target. Principal component analysis is another rigorous method to reduce the number of input variables.14 The method generates a new set of variables, called principal components, that are linear combinations of the original variables:
Dominant Independent Variables.

membership in the set of the weekend moments. It can be explained by fuzzy logic. Fuzzy logic begins where classical logic ends. Fuzzy logic works by assuming the multivalent rather than a bivalent world. Prepositions are not longer regarded as 100% true or 100% false but may fall anywhere on a spectrum between the two extremes. There are four parts in the process of fuzzy inference systems: fuzzification of the input and output variables, fuzzy if-then rules, weighting factors, and defuzzification.17,18
Fuzzication of Input and Output Variables. The first step is to take the inputs and outputs determine the degree to which they belong to each of the appropriate fuzzy sets via membership functions. If Z is the universe and its elements are denoted by z, then a fuzzy set A in Z is defined as a set of ordered pairs:

A = {z , A | z Z } ..........................................................(5)
where A(z) is called the membership function of z in A. The membership function maps each element of Z to a membership value between 0 and 1. A membership function associated with a given fuzzy set maps an input value to its appropriate membership value. The membership function maps each element of Z to a membership value between o and 1. In the first example, the weekend corresponds to the Z set. The days of a weekend to element z and the membership function could be represented by the function in Fig. 4. For variable Friday may corresponds to = 0.65 for the weekend-ness membership function.
Fuzzy If-Then Rules.

y i = e11 x1 + e 21 x 2 + ......ei1 xi ................................................................. (4)

where yi is the principal component i, ei1 is the eigenvector corresponding to the eigenvalues and xs represent the original variables. Each principal component is orthogonal to the other so there is no redundancy. The total variance explained by the original variables is equal to the total variance explained by the principal components. We may use the following rule of thumb: The number of principal components to be used should explain about 90% of total variance. Also, we may use factor analysis to describe the possible relationships among the variables and determine if there is any possibility for grouping variables to reduce the number of independent variables. We use all or these multivariate statistical concepts to make data quality control and determine input dominant variables. Fuzzy Logic The principle for the classical logic is the bivalence; it means any preposition would be either true or false. Aristotle was the person who formulated the Law of the Excluded Middle. Any thing must be either asserted or denied. There is not middle in the principle.16 Prepositions such as half-correct or more or less right are excluded. Human experience suggests something different. The real world is not made up of bivalent blacks and whites, but of grays. There are many examples in the real life; one of those is the concept of weekend. The definition in a dictionary is imprecise: the period from Friday night or Saturday to Monday morning. Fig.4 shows a smoothly varying curve for the fact that parts of Friday partake of the quality of weekend-ness and thus deserve partial

These fuzzy if-then rule statements are expressions to capture the imprecise modes of reasoning that play an essential role in human ability to make decisions in an environment of uncertainty and imprecision. An example that describes a simple fuzzy logic rule is If gamma ray is low, then clay volume is small where gamma ray and clay volume are linguistic variables, low and small are linguistic values or labels that are characterized by membership functions. If there are multiple parts to the antecedent, we must apply fuzzy logic operators (AND, OR, and NOT) and resolve the antecedent to a single number between 0 and 1. In general, the fuzzy operators are defined as follow: AND = min. It takes the minimum membership value from two variables. OR = max. It takes the maximum membership value from two variables. NOT = additive complement. It takes the complement membership value from a variable.
Weighting Factors.

When one rule by itself does not do much good, two or more rules can play off one another. The output of each rule is a fuzzy set and is combined, or aggregated, into a single fuzzy set. The weighting factors of a fuzzy rule

R. SOTO B., ET AL.

SPE 69625

indicate their contribution to the output or decisions of the system. The weighting factors are extracted from human expertise (or experience) and can be adjusted for different situations in a fuzzy inference system.
Defuzzification. The input for the defuzzification is a fuzzy set (the aggregate output fuzzy set) and the output is a crisp value. We use fuzzy logic to develop a rock type model to be used in solving the inverse problem: predict the rock types in uncored intervals/wells.

layers and the number of neurons in each hidden layer. The second is to define a quantitative measure of network performance, called the performance index, which is small when the network performs well and large when the network performs poorly. It represents the calculated mean squared error as the difference between the target output, ykreal, and the network output, ykANN:

e=

1 q

(y
q k =1

real k

y kANN

..........

(7)

Artificial Neural Networks Artificial neural networks, ANN, have inspired in the human brain function. The brain consists of a large number of highly connected elements called neurons. For our purpose these neurons have three principal components: the dendrites, the cell body and the axon (see Fig. 5). The point of contact between an axon of one cell and a dendrite of another cell is called synapses. Fig. 6 shows a simple artificial neuron with more than one input, where the weights, w, correspond to the strength of a synapses, the neuron body is represented by the summation and the transfer function, and the neuron output, y, represents the signal on the axon. The individual inputs x1, x2, xR are weighted by corresponding elements w11, w12, ... w1R of the weight matrix W. The neuron has a bias b, which is summed with the weighted inputs to form the net input: The output, y, depends on the particular transfer function that is chosen. The common transfer functions used in multilayer networks are log-sigmoid and tanh-1. The following expression summarize the neuron output:

where q is the number of data in the training set. The third step is to adjust the network weights and biases in order to reduce the performance index. The most common method used for adjusts the weights and bias is the backpropagation. This method takes the error (difference) from each iteration (training cycle) and uses it to change the weights on the neural network interconnections:

w ij ( t ) =

e + w ij ( t 1) w ij

(8)

y = {tanh 1 ( W1,1 x + b )}

where wij is a weight for the j neuron in the hidden layer i; e is the error from the current training cycle; is the learning rate (a number between 0 and 1 that controls how much the weights can change in each iteration); is the momentum (a constant on the momentum term that uses the previous weight change to keep the errors changing in the right direction); and t reflects the current iteration (t-1 is the previous iteration). We use neural networks to develop a flow zone indicator (FZI) model to be used in solving the inverse problem: calculate FZI in uncored intervals/wells. Case study: Suria and Reforma-Libertad Fields Generalities The Suria and Reforma-Libertad fields belong to the subbasin Apiay-Ariari, located to the southeast of Colombia, approximately to 100 km from Bogot. The producer units in these fields are from the middle Cretaceous (K2, K1 and T1). The K2 unit is the better productive interval, constituted of fluvial sandy bodies. The K1 unit is represented by 33 progressive transgressional subunits and deposits of sand and clay of marine origin form it. As it is typical in this type of environments, the bodies of sand are lateral and vertically heterogeneous. The T1 unit is constituted of marginal and fluvial marine sands of the Oligocene. The main objective of this study is the reservoir characterization of the K1 unit-using core and log data. These fields have 19 and 8 wells respectively and five of them present basic petrophysical analysis (porosity and permeability). Methodology The Methodology followed in this study is the next:

(6)

ANN works very well at solving problems when it is difficult to propose exact mathematical models. Artificial neural networks learn the nature of the dependency between input and output variables through a carefully selected and representative set of training examples. Most artificial neural networks use multi-layer backpropagation architecture to apply knowledge gained from training experiences, which allows the network to make new decisions, classifications, and predictions. The first and the last layer have input and output nodes representing the input and target variables respectively. The interconnections between inputs and outputs are hidden layers (the neurons). A neuron of a hidden layer uses the sum of the weighted outputs of previous layer and non-linear functions that allow the neural network to handle complex, non-linear problems easily and quickly. The number of neurons in each layer and the weights are determined by trials and by optimization 1 (Fig.6). The objective of the neural network is to obtain optimal weights to give a best value for the nodes (the dependent variable) of the output layer. There are three steps involved in development a neural network model. The first step is to define the dominant input variables, the number of hidden

SPE 69625

IMPROVED RESERVOIR PERMEABILITY MODELS FROM FLOW UNITS AND SOFT COMPUTING TECHNIQUES .

A Review of Current Permeability Correlation. Determination of Rock Types from Petrographic Image Analysis Development of Fuzzy Logic Rock Type Model. Hydraulic Flow Unit Classification Using RQI and FZI Concepts Development of Neural Network FZI Model, and Permeability and HFU Prediction

2100 md respectively. Fig 9 represents the prevalent curves of pore size distribution for this rock type. Fig.10 shows the textural characteristics and the distribution of porous spaces.
Rock Type-2.

These rocks are from medium to fine grain size, and possess a homogeneous poral system, defined by intergranular pores, distributed evenly. They present contents of microporosity near to 5% and a low content of kaolinite in the pores. Porosity and permeability values vary from 12 to 19.1% and from 19 to 300 md respectively. These rocks are from medium to fine grain size, with high amounts of pore filling kaolinte. These rocks possess intermediate values of porosity and permeability (9.8 < PHIcore < 16.6 % and 1 < Kcore < 20 md). These low values are owed mainly to the grain size and to its heterogeneous selection.
Rock Type-3.

Conventionally, permeability is estimated from log-derived porosity using equation:

A Review of Current Permeability Correlation.

Log K = a* + b

(9)

Fig.7 shows that relationship between permeability and porosity from core data of Suria and Reforma-Libertad fields. This plot exhibits a high dispersion between these variables. Also, we should point out although the coefficient of determination, R2, is 0.618, the average absolute error is greater than 150%. A residual plot is presented in Fig. 8 to show the inadequacy of that model. We can see in that plot that there is a positive serial correlation and not constant variances. This behavior suggests that permeability depend in this case not only in porosity.
Determination of Rock Types from Petrographic Image Analysis. We use one of standard methods of rock analysis, image analysis. A thin section analysis, using a scanning electronic microscope (SEM), allows identification and determination of total visible porosity (primary intergranular pores, secondary pores and micropores) and pore geometry. Pore geometry controls porosity, permeability, water saturation, irreducible water saturation, electrical properties and capillary pressure. Fundamentally permeability is controlled by pore throat aperture that it is function of grain size, sorting and packing. Therefore, from a visual description of each one of the polished blocks, it is possible to obtain textural attributes, which can serve to classify the porous structures. We carried out a multivariate statistical analysis of pore geometry attributes of 18 samples from the Suria and Reforma-Libertad cores. The analyses are made with image processor equipment, coupled to the SEM. The images observed are obtained using the technique of Back Scatter Electron Detection (BSE). We worked on samples, saturated with epoxy resin and coated with carbon. The poral geometry analysis for this case showed three tendencies of pore distributions. Each group was classified as a rock type. The predominant properties of these rock types are the following ones: Rock Type-1.

The reproduction of rock types defined from petrophysic properties and poral geometry analysis of core samples, to uncored intervals/wells is not a trivial work. This is an inverse process that requires the use of soft computing techniques such as fuzzy logic and neural networks. Before using these techniques, it is necessary to apply a multivariate statistical analysis (correlation matrix, principal components and factor analysis) between well log measurements and core data to find the dominant independent variables. For this particular case, we found that GR and NPHIc represent the behavior of the rock types. We defined the range of each variable from core data (see Table 1) to find the limits of membership functions. Fig. 11 shows the module where the input and ouput variables of fuzzy model are defined. The membership functions were adjusted until the rock type variable was reproduced with a high grade of confidence. Fig.11 shows an example of the rock type prediction.
Development of Fuzzy Logic Rock Type Model. Hydraulic Flow Units Classification using RQI and FZI Concepts.

We apply the concepts of reservoir quality index, RQI and flow unit indicator, FZI to classify core data in hydraulic flow units (HFU) 2,3. Core data is gathered in a single group of data and the corresponding parameters RQI and FZI are calculated, using the following relationships:
RQI = 0.0314 * K core ....................................................... (10) core

FZI = RQI * Z ............................................................. (11a)

or log (RQI) = log(Z ) + log(FZI) ................................... (11b) where Z = core /(1-core). Fig.12 corresponds to log-log plot of RQI vs Z for core data of Suria and Reforma-Libertad fields. We differentiate in this graph, following the Wards algorithm for cluster analysis, three different groups (hydraulic flow units) of data and three straight lines may represent them.

These rocks are from medium grain size, with a variable content of microporosity that not overcome the 5%. They present the best petrophyisic characteristics. Porosity and permeability values vary from 10 to 25% and from 300 to

R. SOTO B., ET AL.

SPE 69625

According to equation 11b, each straight line should have a unitary slope with the intercept equal to mean FZI when Z = 1.0. Also, each HFU has a maximum and minimum FZI values. From equations 10 and 11, we may estimate permeability for each HFU using its mean FZI: core 3 K = 1014 * (FZI mean )2 * 2 (1 core ) ............................. (12)

statistical analysis, fuzzy logic and neural networks to predict rock types, hydraulic flow units and permeability in uncored intervals/wells. 2. The conventional semilog permeability vs. porosity model has an average absolute error greater than 150%. A residual plot of that model shows the inadequacy of that model. There is a positive serial correlation and not constant variance. Permeability, calculated from mean FZI values for each HFU, shows an average absolute error of 67.7%. A residual plot from this approach shows a positive serial correlation and not constant variances. Therefore, this is not a good model to predict permeability in this case meaning that it is not good idea to use mean FZI values ignoring the scatter data for each HFU. We developed a multivariate fuzzy logic model to predict rock types from GR and PHIc for Suria-Reforma fields in Colombia. That fuzzy logic rock type variable was used in conjunction with GR and PHIc to develop a high quality neural network model to predict the FZI. The correlation coefficient for this model is 0.95 with an average absolute error of 6.9%. Permeability is calculated from the neural network FZI and porosity values. The total average absolute error in this case is 18%. A residual from this method shows not apparent serial correlation, the mean error is around zero and therefore the model is valid. Hydraulic flow units are determined from neural network FZI values in uncored intervals/wells. The field results show that our proposed methodology reduces the uncertainty and demonstrates that soft computing techniques is one of the emerging technologies to be taken into account for reservoir engineers in improving reservoir description.

When we compared permeability calculated from the above equation with core permeability, we got an average absolute error of 67.7%. Fig.13 shows a residual plot from this approach where we can see that there is a positive serial correlation and not constant variance. Therefore, this is not a good model to predict permeability in this case. It is explained by the dispersion of data for each HFU in the log-log plot (Fig. 12).
Neural Network FZI Model, and Permeability and HFU Prediction.

3.

4.

We generate a neural network model to predict FZI in uncored intervals/wells. We apply multivariate statistical analysis to determine the dominant input variables. Three input variables were found in this case, GR, NPHIc and Fuzzy Logic Rock Type. The optimal structure of the neural network was composed of two hidden layers of 9 and 11 neurons respectively. Fig. 14 shows a plot between FZI calculated from the neural network model and FZI from core data. The correlation coefficient is 0.95 with an average absolute error of 6.9%. Permeability is obtained applying the equation.12, but in this case we use the value of FZI generated from the neural network model for each sample. Fig. 15 compares permeability calculated from this methodology and core permeability. The total average absolute error in this case is 18%. Checking the model by plotting the residual, Fig. 16, we found that there is not apparent serial correlation, the mean error is around zero and therefore the model is valid. Fig. 17 compares the different permeability models with core permeability in depth function. We can see that the best behavior is followed with the neural net model (KHUFZInn), while the permeability model assuming a mean FZI value (KHUFZIaver) and conventional permeability model (Kconv.) fail in predicting core permeability (Kcore) especially for high values. Finally, HFU is predicted uncored intervals/wells from neural network FZI values. These values are compared with the maximum and minimum FZI values for each HFU, determined before, to check in which range the values fall.

5.

Acknowledgment We express our sincere appreciation to the Instituto Colombiano del Petrleo (Ecopetrol) for providing us the necessary knowledge and computing resources required for this research effort. We also thank to the Reservoir Department of the Gerencia Apiay in Ecopetrol for its support. Nomenclature FZI GR HU Kcore NPHIc PHIcore RQI t = flow Zone Indicator = gamma ray log = hydraulic flow unit = core permeability (md) = corrected neutron log = core porosity (v/v) = reservoir quality index (m) = current iteration

Conclusions 1. This paper suggests a methodology to integrate core and well log measurements. This methodology applies advanced soft computing techniques such as multivariate

SPE 69625

IMPROVED RESERVOIR PERMEABILITY MODELS FROM FLOW UNITS AND SOFT COMPUTING TECHNIQUES .

W x y z Z References

= weights = independent variable = dependent variable = value of individual transform = error signal = learning rate = momemtum = normaliced porosity index.

12.

13.

14. 15.

1. Soto B., R. : "Reservoir Characterization Using Core, Well Log, and Seismic Data and Intelligent Software,"Dissertation, Texas A&M University, College Station, Texas (1998). 2. Amaefule Jude and Altunbay Mehmet: Enhanced Reservoir Description:Using Core and Log Data to identify Hydraulic Flow Units and Predict Permeability in Uncored Intervals/Wells. paper SPE 26436 presented at the 68th Annual Technical Conference and Exhibition. 3-6 Oct-93. 3. Maghsood Abbaszadeh, Hikari Fujii, and Fujio Fujimoto. Permeability Prediction by hydraulic Flow Units- Theory and Applications. Paper SPE 30158 presented at the PetroVietnam Conference 1-3 March -96 4. Timur, A.: An Investigation of Permeability, Porosity, and Residual Water Saturation Relationships for Sandstone Reservoirs, paper presented at the 1968 SPWLA Annual Logging Symposium, New Orleans, 23-26 June. 5. Berg, R.R.: Method for Determining Permeability from Reservoir Rock Properties, Trans., GCAGS (1970) 20, 303. 6. Yao, C.Y. and Holditch, S.A.: Estimating Permeability Profiles Using Core and Log Data, paper SPE 26921 presented at the 1993 Eastern Regional Conference and Exhibition, Pittsburgh, 2-4 November. 7. Amaefule, J.O., Altunbay, M., Tiab, D., Kersey, D.G., and Keelan, D.K.: Enhanced Reservoir Description: Using Core and Log Data to Identify Hydraulic (Flow) Units and Predict Permeability in Uncored Intervals/Wells, paper SPE 26436 presented at the 1993 Annual Technical Conference and Exhibition, Houston, 3-6 October. 8. Soto B., R., Wu, C.H., and Bubela, A. M.: "Infill Drilling Recovery Models for Carbonate Reservoirs -- A Multiple Statistical, Non-Parametric, and Neural Network Approach," paper SPE 57458 presented at the 1999 SPE Eastern Regional Conference and Exhibition, Charleston, WV, October 21-22. 9. Soto B., R., and Holditch, S.A..: "Development of Reservoir Characterization Models Using Core, Well Log, and 3D Seismic Data and Intelligent Software," paper SPE 57457 presented at the 1999 SPE Eastern Regional Conference and Exhibition, Charleston, WV, October 21-22. 10. Nikravesh, M., Kovscek, A.R., Johnston, R.M., and Patsek, T.W.: Prediction of Formation Damage During Fluid Injection Into Fractured, Low Permeability Reservoirs via Neural Networks, paper SPE 31103 presented at the 1996 SPE Formation Damage Symposium, Lafayette, LA, 16-18 February. 11. Garg, A., Kovscek, A.R., Nikravesh, M., Castanier, L.M., and Patzek, T.W.: CT Scan and Neural Network Technology for Construction of Detailed Distribution of Residual Oil During Waterflooding, paper SPE 35737

16. 17. 18.

presented at the 1996 SPE Western Regional Meeting, Anchorage, AK, 22-24 May. Juniardi, I.R. and Ershaghi, I.: Complexities of Using Neural Network in Well Test Analysis of Faulted Reservoirs, paper SPE 26106 presented at the 1996 SPE Western Regional Meeting, Anchorage, AK, 22-24 May. Zhou, C.D., Wu, X.L., and Cheng, J.A.: Determining Reservoir Properties in Reservoir Studies Using a Fuzzy Neural Network, paper SPE 26430 presented at the 1993 SPE Annual Tech. Conference and Exhibition, Houston, TX, 8-12 October . Johnson, R.A., and Wichern, D.W.: Applied Multivariate Statistical Analysis, Fourth Edition, Printice-Hall, Inc., Upper Saddle River, NJ (1998). Jhon-Se Lim, Joe M. Kang, and Jungwham Kim,Multivariate Statistical Analysis for Automatic Electrofacies determination from well log measurements, paper SPE 38028, presented at the 1997 SPE Asia Pacific Oil and Gas Conference. Brown Malcom.:Fuzziness in the Field. Review Oct-Dec 1998. Fuzzy Logic Tollbox. Users Guide. Version 2 MATLAB. Sept 2000. Xiong Xongjie and Holditch S.A. An Investigation Into the Application of Fuzzy Logic to Well Stimulation Treatmnt Design. Paper SPE 27672 present at the Permian Basin Oil and Gas Recovery Conference. March 16-18-94.

Table.1 Range of Variables for each Rock Type.

Variable
Min. GR NPHIc PHICORE KCORE FZI
18.32 0.113 0.126 479.6 9.34

SUMMARY TABLE Rock 1 Rock 2

Max.
79.19 0.239 0.216 2977.0 17.75

Rock 3
Max. Aver.
107.6 0.200 0.158 5.76 1.17 63.3 0.110 0.100 1.77 0.70 54.9 0.130 0.149 156.7 4.15 34.7 0.058 0.052 0.00 0.07

Aver.
38.74 0.181 0.164 1131.6 12.79

Min.
21.5 0.060 0.034 0.2 1.35

Max. Aver. Min.

114.7 0.192 0.217 816.3 7.63

Table.2- Absolute Average Error for each Permeability Model.

PERMEABILITY SUMMARY TABLE Average Absolute Error Kconv. KHUFZIaver. KHUFZInn. > 150 % .= 67.7% .= 18%

R. SOTO B., ET AL.

SPE 69625

Fig. 4. Representation of Weekend-ness Membership Function. Fig. 1. Q-Q Plot for Gamma Ray Variable Rocks 1, 2 and 3.

Fig. 2. Q-Q Plot for Gamma Ray Variable Rocks type 3.

Fig. 5 A Simple Neuron and Communication between Neurons.

Fig. 3. A 95% Confidence Ellipse for Checking of any Suspicious Observations (Outliers).

Fig. 6 Neural Network Architecture.

SPE 69625

IMPROVED RESERVOIR PERMEABILITY MODELS FROM FLOW UNITS AND SOFT COMPUTING TECHNIQUES .

Porosity/Permeability Relationship
10000 1000 Permeability(mD)

ERROR > 150%

100 10 1 0.1 0.01 0.001 Porosity(v/v) 0 0.05 0.1 0.15 0.2 0.25
55.801x

y = 0.0145e 2 R = 0.6187

Fig.7- Log Permeability versus Porosity, Suria and ReformaLibertad Samples.

Fig.10- Microghraph of Rock Type 1 (Suria and ReformaLibertad Fields).

RESIDUAL PLOT
1000 Kconv 500

Kconv-KCORE

ERROR > 150%

0
0 500 1000 1500 2000 2500 3000

-500

-1000

KCORE(mD)

Fig.8 Residual Plot Between Core Permeability (Kcore) and Conventional Permeability (Kconv).
ROCK TYPE 1
30%

Fig.11- Module of Definition of Input and Output Variables for Fuzzy Logic Type Rock Model.
Relation Z versus RQI

10.000
25%

20%

POROSITY(%)

0.010
RQI
15%

0.100

1.000 1.000 0.100

10%

0.010
5%

FZI 0.001

0% 1 10 100 1000

(normalized porosity)

PORE SIZE(micr)

Fig.9- Pore Size Distribution, Rock Type 1 (Suria and Reforma-Libertad Fields).

Fig.12- Log-Log Plot of Z versus RQI for Flow Unit Identification.

10

R. SOTO B., ET AL.

SPE 69625

RESIDUAL PLOT
1000 KHUFZIaver

RESIDUAL PLOT
1000
KHUFZInn

KHUFZInn-KCORE

KHUFZIaver-KCORE

500 ERROR = 67.7% 0 0 -500 500 1000 1500 2000 2500 3000

500 ERROR = 18% 0 0 -500 500 1000 1500 2000 2500 3000

-1000 KCORE(mD)

-1000 KCORE(mD)

Fig. 13 Residual Plot between Core Permeability (Kcore) and Permeability Calculated Using mean FZI values (KHUFZIaver).
NNFZI/FZI RELATIONSHIP
25

Fig.16 Residual Plot between Core Permeability (KCORE) and Permeability Calculated Using Neural Network FZI Model (KHUFZInn).
COMPARISON OF PERMABILITY MODELS
3000

PERMEABILITY MODELS(mD)

NNFZI(neural network)

20 15 10 y = 0.9525x + 0.0034 R2 = 0.95

2500 2000 1500 1000 500 0

Kconv KHUFZIaver KHUFZInn KCORE

ERROR = 6.9%
5 0 0 5 10 15 20 25

FZI(cores)

Fig. 14 Correlation Between Core FZI and NNFZI.

CORE PERMEABILITY Vs NEURAL NETWORK PERMEABILITY
3500 3000

Fig.17 Comparison between Core Permeability(Kcore) and Permeability Models: Conventional Permeability Model (Kconv), Permeability Using mean FZI values (KHUFZIaver), and Permeability Using Neural Network FZI Model (KHUFZInn).

KHUFZInn(mD)

2500 2000 1500 1000 500 0 0 500 1000 1500 2000 2500 3000 3500

ERROR = 18%

KCORE(mD)

Fig.15 Comparison Between Core permeability (Kcore) and neural network permeability.

68 10 7 20 10 4 50 10 7 49 10 5 94 10 8 47 10 2 29 10 7 29 10 5 94 10 6 78 10 2 78 10 5 81 10 9 21 10 8 68 1

10

CORE DEPTH(ft)