SciELO - Scientific Electronic Library Online

 
vol.1 número5PRESSURE CONTROL IN VACUUM TOWERS índice de autoresíndice de assuntospesquisa de artigos
Home Pagelista alfabética de periódicos  

Serviços Personalizados

Journal

Artigo

Indicadores

Links relacionados

  • Em processo de indexaçãoCitado por Google
  • Não possue artigos similaresSimilares em SciELO
  • Em processo de indexaçãoSimilares em Google

Compartilhar


CT&F - Ciencia, Tecnología y Futuro

versão impressa ISSN 0122-5383versão On-line ISSN 2382-4581

C.T.F Cienc. Tecnol. Futuro v.1 n.5 Bucaramanga jan./dez. 1999

 

DEVELOPMENT OF INFILL DRILLING RECOVERY MODELS FOR CARBONATE RESERVOIRS USING NEURAL NETWORKS AND MULTIVARIATE STATISTICAL AS A NOVEL METHOD

R. SOTO*2, CH. H., WU1, and A. M. BUBELA1


1Texas A & M University, College Station, Texas 77843-3116, USA
2Ecopetrol - Instituto Colombiano del Petróleo, A.A. 4185 Bucaramanga, Santander, Colombia

E-mail: byron.perez@ieee.org

*A quien debe ser enviada la correspondencia


Abstract

This work introduces a novel methodology to improve reservoir characterization models. In this methodology we integrated multivariate statistical analyses, and neural network models for forecasting the infill drilling ultimate oil recovery from reservoirs in San Andres and Clearfork carbonate formations in West Texas. Development of the oil recovery forecast models help us to understand the relative importance of dominant reservoir characteristics and operational variables, reproduce recoveries for units included in the database, forecast recoveries for possible new units in similar geological setting, and make operational (infill drilling) decisions. The variety of applications demands the creation of multiple recovery forecast models. We have developed intelligent software (Soto, 1998), Oilfield Intelligence (Ol), as an engineering tool to improve the characterization of oil and gas reservoirs. Ol integrates neural networks and multivariate statistical analysis. It is composed of five main subsystems: data input, preprocessing, architecture design, graphic design, and inference engine modules. One of the challenges in this research was to identify the dominant and the optimum number of independent variables. The variables include porosity, permeability, water saturation, depth, area, net thickness, gross thickness, formation volume factor, pressure, viscosity, API gravity, number of wells in initial waterflooding, number of wells for primary recovery, number of infill wells over the initial waterflooding, PRUR, IWUR, and IDUR. Multivariate principal component analysis is used to identify the dominant and the optimum number of independent variables. We compared the results from neural network models with the non-parametric approach. The advantage of the non-parametric regression is that it is easy to use. The disadvantage is that it retains a large variance of forecast results for a particular data set. We also used neural network concepts to develop recovery models. The neural network infill drilling recovery model is capable of forecasting the oil recovery with less error variance compared with non-parametric, fuzzy logic and regression models.

Keywords: neural networks, carbonate reservoirs, recovery models, infill drilling.


Resumen

Este trabajo introduce una metodología novedosa para mejorar los modelos de caracterización de yacimientos. En esta investigación se usaron técnicas de estadística multivariada y redes neuronales para desarrollar modelos de predicción de los recobros primarios de aceite (PRUR), recobros de aceite al inicio de la inyección de agua (IWUR) y recobros de aceite debido a la perforación de pozos de relleno (IDUR) en yacimientos de carbonatas localizados en el este de Texas. Los modelos desarrollados fueron comparados con los modelos de regresión no-lineal y con los de regresión no-paramétrica. Uno de los desafíos en esta investigación fue identificar las variables independientes dominantes y el número óptimo de estas. Para ello se desarrolló un sistema inteligente (Soto, 1998), Oilfield Intelligence (Ol), que integra conceptos de componentes principales, análisis de factores y redes neuronales. Ol está compuesto por cinco subsistemas: carga y preprocesamiento de los datos, diseño de la arquitectura de la red neuronal, diseño gráfico y una máquina de inferencia. El análisis multivariado de componentes principales permite resolver el problema de dimensionalidad. Cuántas y cuáles variables deberían usarse en la obtención de cada modelo. Después se utilizaron las redes neuronales para desarrollar modelos capaces de predecir los recobros primarios, de inyección de agua y debido a la perforación de pozos de relleno en las formaciones de carbonato de San Andrés y Clearfork en el este de Texas. Los coeficientes de correlación son del orden del 99% con errores absolutos no mayores del 3% comparados con coeficientes de correlación del orden de 0.91 y errores absolutos alrededor del 27% de otros modelos publicados internacionalmente en los últimos 1 5 años. Las variables consideradas en esta investigación fueron porosidad, permeabilidad, saturación de agua, profundidad, área, espesor total, espesor neto, factor volumétrico de formación, presión, viscosidad, gravedad API, número de pozos al inicio de la inyección de agua, número de pozos para la recuperación primaria, número de pozos de relleno al inicio de la inyección de agua, PRUR, IWUR, e IDUR. Obviamente el desarrollo de un modelo en redes neuronales que represente con alta precisión los datos requiere experiencia del ingeniero para realizar un control de calidad de los datos, determinar las variables dominantes y optimizar la estructura o topología de la red neuronal.

Palabras clave: redes neuronales, yacimientos de carbonatas, modelos de recuperación, perforación de relleno.


INTRODUCTION

The amount of oil that can be recovered from an oil rervoir is dependent on the reservoir characteristics, recovery method, number of wells, and operations efficiency. Waterflooding is often used as a secondary recovery process after the reservoir has been produced during primary recovery. Since most reservoirs are heterogeneous, infill drilling after an initial water flood permits production of oil from parts of the reservoir that might otherwise have been bypassed (Wu et al. 1992; Wu et al., 1988; Shao et al., 1994a; French et al., 1991). Researchers and engineers have shown that the infill drilling in the West Texas carbonate reservoirs has indeed accelerated and increased oil recovery (Lu et al., 1993, 1994). However, the dominant mechanisms and parameters (independent variables) that affect the oil recovery are not fully identified or may be subjected to varying degrees of uncertainty. To forecast or predict infill drilling recovery efficiency of an individual unit or a reservoir is a difficult task for most reservoir-engineering professionals.

Many approaches are used to evaluate infill drilling recovery efficiency. One is based on unit recovery efficiency and reservoir analysis, plus the reservoir engineer´s experience and intuition. A more elaborate approach is the use of a "reservoir simulator", which requires detailed and yet uncertain input reservoir and production data. The approach we take here is statistical. It is based on an oil recovery efficiency database developed for specific producing formations in a particular geological basin.

The field data from waterflood units in the Permian Basin were gathered and used to develop the oil recovery forecast models. The database includes reservoir and production data of 21 units in the San Andres formation, and 23 units in the Clearfork formation. Estimation of ultimate and incremental infill drilling recovery has been a difficult task because of the limited data and the uncertainties in the independent variables (Malik et al., 1993; Shao et al., 1994b; Wu et al., 1993, Wu et al., 1997). Many attempts have been made to develop the infill drilling oil recovery forecast models with reference to reservoir rock and fluid properties. Nonlinear regression, statistical analysis, and fuzzy logic were used to analyze the oil recovery data. While these reported models have improved over the years, they are not entirely satisfactory due to the inexact nature of the data set and the inherent limitations in the models themselves.

With the development of linear regression (Wu et al., 1998; Shao et al., 1994a) or statistical analysis (French et al., 1991) models of infill drilling ultimate recovery (IDUR), a substantial effort was put into determining which independent variables are the most important and what is the optimum number of the independent variables. Some of these independent variables go beyond the basic reservoir and field properties. Additional parameters such as the primary ultimate recovery (PRUR) and the initial waterflood ultimate recovery (IWUR) were needed to succesfully forecast IDUR (Lu, 1993).

In this work non-parametric regression analysis neural network modeling are used to develop forecast models that improve the degree of consistency and accuracy. The non-parametric approach is based on the work of Breiman and Friedman (1985). Xue and Datta-Gupta (1996) applied this approach to integrate seismic data in reservoir characterization. While the non-parametric regression analysis provided better identification of the dominant independent variables and more consistent forecast, it did not improve the variance of forecast results.

The application of neural networks for modeling non- linear systems has been improved substantially in recent years (Nikravesh et al., 1996; Bomberger et al., 1996; Rogers and Dowla, 1994; Al-Kaabi et al., 1990; Azimi-Sajadi and Liou, 1989; Johnson and Wichern, 1998). How ever, one of the problems that still had to be addressed was the determination of the dominant variables, and the optimum number of independent variables. Some intuitive insight and functional knowledge of the physical behavior of the system turned out to be helpful for identifying the dominant independent variables. A nonparametric regression analysis and the principal components and factor analysis of multivariate statistical anal ysis were used to identify the dominant independent variables so we could develop more efficient and realistic neural-networks models.

DATA SET

The input data are shown in Tables 1, 2, 3, and 4. Tables 1 and 2 list the reservoir and operational properties of the San Andres units. The dependent variables are primary ultimate oil recovery (PRUR), initial waterflood ultimate recovery (IWUR) and infill drilling ultimate recovery (IDUR). The labels and units of each variable are referred to the Nomenclature. During the data analysis and model development, it is found that IWUR is strongly dependent on PRUR; and IDUR on PRUR and IWUR. The sequential dependency complicates the development of dependable IDUR models. Table 3 and 4 list the properties of Clearfork units.

A REVIEW OF THE INFILL DRILLING RECOVERY MODELS

The updated non-linear standard regression forecast models for primary ultimate oil recovery (PRUR), initial waterflooding ultimate oil recovery (IWUR), and infill drilling ultimate oil recovery (IDUR) of San Andres and Clearfork units are summarized in Table 5. For each model, we included the independent variables, the coefficient of determination and the average absolute error. Figure 1 shows a comparison of actual and predicted PRUR. Apparently, the correlation between predicted values and actual values is good but the average absolute error is about 23.46%. Figures 2, 3 and 4 show residual plots for the statistical model to predict PRUR, IWUR, and IDUR. The residual plots do not have a constant variance and the upward trend suggests that the models may need additional terms. From these plots and the average absolute errors, we could conclude that it was necessary to search other modeling techniques.

NON-PARAMETRIC REGRESSION APPROACH FOR ESTIMATING OPTIMAL TRANSFORMATIONS FOR MULTIPLE REGRESSIONS

Non-parametric regression is one of the novel approaches to constructing a suitable model description from available information. It is developed to alleviate the problem of parametric regression that often leads to erroneous results caused by the mismatch between assumed model structure and the physical relationships of the actual data. In non-parametric regression we do not fix a priori the form of the dependency of the dependent variable on the independent variables. In fact, one of the main results of non-parametric regression is the form of the relationship.

Non-parametric regression is intended to build a model in the form,

where, the inverse transformation, f0-1 ,and the transform sum of the independent variables, z0 ,are selected to maximize the correlation between the right- hand and left-hand sides of the relation:

subject to some constraints. The data transforms is calculated:

In this case the symbol fn(xn) does not necessarily mean a certain algebraic expression. It is rather a relationship defined point-wise. The method of alternating conditional expectations (ACE) (Breiman and Friedman 1985), constructs and modifies the individual transformations in order to maximize the correlation in the transformed space. Certain trivial constraints (zero mean and unit variance for the individual transformations) assure that the solution is almost unique. To make the ACE algorithm really work, however, one has to imply some kind of restriction on the smoothness of the individual transformations, and this is done somewhat hidden, the way a certain "smoother" is used to construct and improve the transformations.

One of the great advantages of non-parametric regression is that it provides an insight into the influence of the individual variables. The shape of the point-to- point transformation is very informative, and the range of the transformed variable zi tells a lot about the relative significance of the independent variables.

We used a dummy variable called BASIN with values 1 for San Andres and 2 for Clearfork. The coefficients of determination and the average absolute errors for PRUR, IWUR, and IDUR are 0.9786, 0.933 and 0.9436, and 21.2%, 28.1% and 29.4 %, respectively. figure 5 shows a comparison of actual and predicted PRUR for non-parametric model using GRACE. Figure 6 shows the residual plot for IDUR to check for adequacy of each model. We can see from this plot that the errors do not have zero mean, neither constant variances, and the model underestimates the IDUR as the actual IDUR increases.

THE OILFIELD INTELLIGENCE

The neural network simulator developed in this research is called Oilfield Intelligence (OI) (Soto, 1998). The neural network simulator was built in MATLAB, a high-performance language for technical computing that integrates computation, visualization, and programming in an easy-to-use environment.

OI is composed of five main subsystems: Loading data, preprocessing, architecture design, graphic design, and inference engine modules. We wrote more than 1,200 lines of programming as M-files using MATLAB as a platform. Figure 7 illustrates the architecture OI and Figure 8 shows the graphical user interface (GUI) with each of these modules.

THE DOMINANT INDEPENDENT VARIABLES

A multivariate statistical analysis was also performed to determine the dominant independent variables with reference to that investigated by the non-parametric regression analysis. Table 6 shows the results of a principal component analysis for PRUR. As can be seen, seven principal components could explain about 90% of the total variance of the data. To describe the possible relationships among the variables and determine if there is any possibility for grouping variables, we used the concept of factor analysis (Johnson and Wichern 1998). Table 7 shows an output of the rotated factor loading. The first factor grouped the primary ultimate oil recovery (PRUR) with the productive area and the number of primary recovery wells (NOPW) as indicated by the loading values. The first factor shows the dependence of the primary ultimate oil recovery on the productive area and the number of wells for primary recovery. According to this factor analysis, we decided to use area, because of the highest coefficient in factor1, as independent variable to predict PRUR. The second factor grouped BASIN, DEPTH and initial reservoir pressure (PRESS). For this group, we selected BASIN as independent variable because it was easier for the neural network to recognize that pattern. The third factor grouped the API gravity and FVF. The fourth factor shows that gross thickness (GROSS) has the highest loading factor. Factors five and six are explained by water saturation (SW) and porosity (POR) variables respectively. Similar principal component and factor analysis was performed for IWUR and IDUR.

Results of the principal component and factor analysis indicate that the dominant independent variables for PRUR are: productive area (AREA), BASIN, API gravity, FVF, gross thickness (GROSS), and initial water saturation (SW). The dominant independent variables for IWUR are: productive area, the number of initial waterflood wells (NOWW), DEPTH, the number of primary wells (NOPW), API, FVF, net pay (NET), porosity (POR), permeability (PERM), and viscosity (VIS). The dominant independent variables for IDUR are: BASIN, productive area (AREA), and the number of infill wells (NOIW), PRESS, API, FVF, NET, VIS, SW, and POR. Those dominant independent variables were used to develop neural network infill drilling recovery models.

NEURAL NETWORKS

A neural network is a series of layers with nodes and weights that represent complex relationships among input and output variables. The first layer has input nodes representing the input variables (independent variables) specified by the problem. The node of a hidden layer uses the sum of the weighted outputs of previous layer and sigmoid function to provide output for the nodes in the subsequent hidden layer. The number of nodes in each layer and the weights are determined by trials and by optimization. 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.

The advantages of the neural network approach are several. It does not require an explicit functional relationship between the input and output variables. It can be trained from past available data to learn and approximate the nonlinear relationships to any degree of accuracy. It is applicable to multivariate systems. One of the significant drawbacks of the neural network approach is that the input nodes must be specified a priori. The type of input variables and the optimal number of input variables can not easily be determined from neural network analysis.

Performance learning is one of the most important steps where network parameters are adjusted in an effort to optimize the performance of the network. Two steps are involved in this optimization process. The first step is to define a quantitative measure of network performance, performance index. It represents a global error (E) of the neural network defined as,

Where the inner summation is over all nodes in the output layer, and the outer sum is over the number of the training set. During training, the size of error generally decreases until it reaches a threshold level.

The second step of the optimization process is to search the parameter space to reduce the performance index. A steepest descent algorithm is often used for the optimization. Mathematically, the weights (W) are adjusted as follows:

Where the inner summation is over all nodes in the output layer, and the outer sum is over the number of the training set. During training, the size of error generally decreases until it reaches a threshold level.

The second step of the optimization process is to search the parameter space to reduce the performance index. A steepest descent algorithm is often used for the optimization. Mathematically, the weights (W) are adjusted as follows:

Where η is learning rate and o is the output value from hidden nodes and δ is the error signal term produced by the nodes in the hidden or output layer,

Before training the network, the learning rate (η) for the network must be specified. The new weights are then used to calculate the new output. The procedure is repeated until a tolerance is satisfied.

We used a backpropagation algorithm with the Levenberg-Marquardt procedure as an optimization method for convergence was used. The 44-sample data set was divided into three subsets for training set (72%), validation (14%) and testing (14%). A post-training evaluation of the performance of the trained neural networks is carried out by calculation of the errors for the training, validation, and testing data sets.

The final topology of the neural network for prediction of PRUR has 7 neurons (independent variables) in the input layer and two hidden layers with 12 and 10 neurons, respectively. For IWUR, the topology has 9 neurons in the input layer and two hidden layers with 10 and 6 neurons respectively. The topology of the neural network for IDUR has 9 neurons in the input layer, and two hidden layers with 10 and 8 neurons, respectively. After the neural networks were "trained" the weight and bias vectors were incorporated into Fortran-90 and Visual Basic interfaces so that the results could be used in a practical manner.

The neural network model showed very good performance for prediction of PRUR, IWUR, and IDUR.

The comparison of calculated and measured PRUR, IWUR and IDUR are presented in Figures 9 to 11.

The coefficients of determination and the average absolute errors for PRUR, IWUR, and IDUR are 0.998, 0.992 and 0.9995, and 1.0%, 2.1% and 3.3%, respectively. Using the same scales of the residual plots for non-parametric regression models, we plotted residual plots for the neural network models of PRUR, IWUR, and IDUR Figure 10 shows an example). Checking the adequacy of each model we can see from this plot that the errors do have zero mean and constant variance.

The dominant independent variables identified for each model are used to develop the neural network oil recovery models. A series of sensitivity analysis with different neural network topologies is performed to develop the best neural network models. The approach helped eliminate over-fitting and meaningless dependency of certain independent variables. For the sensitivity analysis of variable dependency, we made a series of runs for each basin (San Andres and Clearfork) by varying the value of each independent variable while keeping other independent variables at individual mean. As an example, Figure 12 shows the dependency of calculated IDUR on the productive area. The monotonely increasing relationship indicates physically meaningful dependency of the variables. Figure 13 shows the dependency of the calculated IDUR on the number of infill wells. Similar sensitivity analysis was made for other independent variables. Table 8 shows a summary of the oil recovery forecast model performance.

CONCLUSIONS

  • The correlation coefficients of the non-linear regression models for predicting the infill ultimate oil recovery for both San Andres and Clearfork carbonate formations in West Texas apparently are good but the average absolute error is about 23.46%
  • One of the significant constraints for the model development is the limited number of field data that are inexact and often exhibit uncertain relationships. Principal components and factor analysis help understand the relative importance of dominant reservoir characteristics and operational variables to improve the modeling.
  • The advantage of the non-parametric regression is that it is easy to use and can quickly provide results that reveal the dominant independent variables and relative characteristics of the relationships. The disadvantage is retaining a large variance of forecast results for a particular data set. The average absolute errors for PRUR, IWUR, and IDUR are 21.2%, 28.1% and 29.4%, respectively. The residual plots showed that the errors do not have zero mean, nor constant variances.
  • Multivariate principal component and factor analyses were employed to develop an effective neural network. The neural network infill drilling recovery model is capable of forecasting the oil recovery with less error variance. The average absolute errors for PRUR, IWUR, and IDUR are 1.0%, 2.1% and 3.3% respectively. The residual plots showed that the errors do have zero mean and constant variances.
  • The novel methodology applied in this research can be used to get better models for a reservoir characterization.

ACKNOWLEDGMENT

This material is based in part upon work supported by the Texas Advanced Technology Program under Grant No. ATP-036327-066 1998. The Instituto Colombiano del Petroleo support for part of the works is also appreciated. We would like to acknowledge Peter Valko's contribution to the development and application of multiple forecast models.


REFERENCES

Al-Kaabi, A.U., McVay, D.A. and Lee, W. J., 1990. "Using an Expert System to Identify a Well-Test Interpretation Model", J. Pet. Tech. (May): 654 - 661.        [ Links ]

Azimi-Sajadi, M. and Liou, R. J., 1989. "Fast Learning Process of Multilayer Neural Networks Using Recursive Least Squares Method", IEEE Trans. On Sig. Proc., 40 (2): 446 - 449.        [ Links ]

Bomberger, J. D., Seborg, D. E., Lightbody, G., and Irwin, G. W., 1996. "Experimental Evaluation of Neural Nonlinear Modeling", presented at the Chemical Process Control V, Tahoe City, CA. (26 January).        [ Links ]

Breiman, L. and Friedman, J. H., 1985. "Estimating Optimal Transformations for Multiple Regression and Correlation", J. of the American Statistical Associations (September): 580.        [ Links ]

French, R. L., Brimhall, R. M., and Wu, C. H., 1991. "A Statistical and Economic Analysis of Incremental Waterflood Infill Drilling Recoveries in West Texas Carbonate Reservoirs," SPE 22624 presented at the SPE Annual Technical Conferences and Exhibition, Dallas (Oct. 6 -9).        [ Links ]

Johnson, R. A. and Wichern, D. W. 1998. Applied Multivariate Statistical Analysis, Fourth Edition, Printice-Hall, Inc., Upper Saddle River, NJ.        [ Links ]

Lu, G. F., Brimhall, R. M. and Wu, C. H., 1993. "Geographical Distribution and Forecast Models of Infill Drilling Oil Recovery for Permian Basin Carbonate Reservoir", SPE 16503 poster session at the nnual Technical Conference and Exhibition, Houston (Oct. 3 - 6).        [ Links ]

Lu, G. F., Jagoe, B. L. and Wu, C. H., 1994. "Technical Factors Useful for Screening Reservoirs for Water-flood Infill Drilling", SPE 27660 (poster) presented at the Permian Basin Oil & Gas Conference, Midland (March 16 - 18).        [ Links ]

Malik, Z. A., Silva, B. A., Brimhall, R. M. and Wu, C. H., 1993. "An Integrated Approach to Characterize Low Permeability Reservoirs Connectivity to Optimize Waterflood Infill Drilling", SPE 25853 presented at the Joint Rocky Mountain Regional and Low Permeability Reservoirs Symposium, Denver (April 26 - 28).        [ Links ]

Nikravesh, M., Kovscek, A. R., Johnston, R. M. and Patsek, T. W., 1996. "Prediction of Formation Damage During Fluid Injection Into Fractured, Low Permeability Reservoirs via Neural Networks", paper SPE 31103 presented at the SPE Formation Damage Symposium, Lafayette, L A. (February 16 - 18).        [ Links ]

Nikravesh, M., Kovscek, A. R., Patzek, T.W. and Soroush, M., 1996. "Identification and Control of Industrial Scale Processes via Neural Networks", presented at the Chemical Process Control V, Tahoe City, CA. (January 26).        [ Links ]

Rogers, L.L. and Dowla, F. U., 1994. "Optimization of Groundwater Remediation Using Artificial Neural Networks with Parallel Solute Transport Modeling", Water Res. Res., 30 (2): 457 - 481.        [ Links ]

Shao, H., Brimhall, R. M., Ahr, W. M. and Wu, C. H., 1994a. "Integrated Recovery Efficiency Forecast Models for San Andres Reservoirs of Central Basin Platform and Northern Shelf, West Texas", SPE 27697 (poster) presented at the Permian Basin Oil & Gas Conference, Midland (March 16 - 18).        [ Links ]

Shao, H., Brimhall, R. M., Ahr, W. M. and Wu, C. H., 1994b. "An Application of Idealized Carbonate Depositional Sequences for Reservoir Characterization", SPE 28458 (poster) presented at the Annual Technical Conference and Exhibition, New Orleans (Sept. 25-28).        [ Links ]

Soto B., R., 1998. "Reservoir Characterization Using Core, Well Log, and Seismic Data and Intelligent Software", Dissertation, Texas A & M University, College Station, Texas.        [ Links ]

Wu, C. H., Lu, G. F., Xue, G. P. and Brimhall, R. M., 1993. "Empirical Oil Recovery Forecast Models for Waterflood Infill Drilling in West Texas Carbonate Reservoirs", presented at the Southwestern Petroleum Short Course, Texas Tech. University, Lubbock, Texas (April 21 - 22).        [ Links ]

Wu, C. H., Lu, G. F., Gillespie, W. and Yen, J., 1997. "Statistical and Fuzzy Infill Drilling Models for Carbonate Reservoirs", paper SPE 37728 presented at the SPE Middle East Oil Show & Conference, Bahrain (March 15 - 18).        [ Links ]

Wu, C.H., 1992. "Infill Drilling Use Improves Waterflood Oil Recovery", The American Oil & Gas Reporter (61 March).        [ Links ]

Wu, C. H., Laughlin, B. A. and Jardon, M., 1988. "Impact of Infill Drilling on Waterflood Recovery: West Texas Carbonate Reservoirs", SPE 17286 presented at the SPE Permian Basin Oil and Gas Recovery Conference, Midland (March 10-11).        [ Links ]

Xue, G. and Datta-Gupta, A., 1996. "A New Approach to Seismic Data Integration During Reservoir Characterization Using Optimal Non-Parametric Transformations", paper SPE 36500 presented at the SPE Annual Technical Conference, Denver (October 6 - 9).        [ Links ]

Xue, G., Datta-Gupta, A., Valkó, P. and Blasingame, T., 1996. "Optimal Transformations for Multiple Regression: Application to Permeability Estimation From Well Logs", paper SPE 36500 presented at SPE Improved Oil Recovery Symposium, Tulsa (April 21).        [ Links ]

Creative Commons License Todo o conteúdo deste periódico, exceto onde está identificado, está licenciado sob uma Licença Creative Commons