--> Predicting Log Properties from Seismic Data Using Abductive Networks
[First Hit]

Datapages, Inc.Print this page

Click to view page image in pdf format.

AAPG Bulletin, Vol. 90 (2006), Program Abstracts (Digital)

7th Middle East Geosciences Conference and Exhibition
Manama, Bahrain
March 27-29, 2006

ABSTRACT: Predicting Log Properties from Previous HitSeismicNext Hit Data Using Abductive Networks

Osama A. Ahmed1, Radwan Abdel-Aal2, and Husam AlMustafa3
1 Applied Electrical Engineering, KFUPM, Hail Community Coleege, KFUPM, Hail, Saudi Arabia, phone: 0564203171, [email protected]
2 Computer Engineering Dept, KFUPM, KFYPM, Dhahran, 31261, Saudi Arabia
3 ARAMCO, Dhahran, Saudi Arabia

In this study, abductive network is used to predict reservoir log properties from Previous HitseismicNext Hit attributes. Statistical approaches have been used to model the relationship between the Previous HitseismicNext Hit data and the reservoir parameters. The idea of using multiple Previous HitseismicNext Hit attributes to predict log properties has been widely used and several case histories have been reported in the literature using multi-linear stepwise regression and neural networks. The input to any statistical method is a series of attributes extracted from the Previous HitseismicNext Hit data. There is, however, a huge number of attributes that can be extracted form the Previous HitseismicNext Hit dataý. Therefore, an efficient subset of this attributes has to be selected before prediction. Exhaustive search of all Previous HitattributeNext Hit combinations is computationally infeasible. As a solution, linear stepwise regression has been proposed which is based on linear relationships between Previous HitattributeNext Hit combinations and log data. Therefore it is suitable for linear regression. For non linear regression such as neural networks an Previous HitattributeNext Hit selection method that embodies the nonlinearity between Previous HitattributeNext Hit combinations and log data is desirable. Abductive Networks should in many ways help in this regard: 1. Abductive Networks can automatically select a statistically representative subset of optimum predictors from the available set of Previous HitseismicNext Hit attributes. 2. Abductive Networks are nonlinear predictors which are proven to outperform linear predictors ý. 3. Unlike various neural network paradigms, Abductive Networks can provide a closed form analytical relationship between the selected Previous HitseismicTop attributes and the modeled parameter; this can help in fully understanding the geographical structure of the area.

 

Copyright © 2006. The American Association of Petroleum Geologists. All Rights Reserved.