Neural Networks and Seismic Reservoir Characterization

Dr. Hongliu Zeng
Bureau of Economic Geology
John A.& Katherine G. Jackson School of Geosciences
The University of Texas at Austin

Abstract
Artificial neural networks (ANNs) are computer programs designed to simulate the way a simple biological nervous system is thought to operate. The networks have the capacity to learn the relationships among complex data that are very difficult, or even impossible, to express explicitly as mathematical equations. Many problems in geoscience are complicated, fuzzy, or simply un-resolvable using present methods. Application of ANNs may provide alternative solutions that are accurate and cost effective.

In seismic-guided reservoir characterization, the most difficult task is to understand and formulate all factors that significantly influence how a seismic wave responds to a rock. Assumptions and simplifications often result in unrealistic and erroneous solutions. Without these limitations, however, a neural network can find the relationship between seismic attributes and reservoir properties in a self-learning and nonlinear way, providing training data containing all relevant information. Examples from certain Bureau research projects (for example, SGR, Sacroc, and Fullerton) show that ANNs significantly improved the mapping of reservoir properties (for example, Volume shale, porosity, and P-velocity) from 3-D seismic data, in both clastic and carbonate sequences.

For a successful application of neural-network methods, a clear understanding of the underlying physics of the problem is essential. Seismic data must be sensitive enough to the rock property of interest; and fundamental ambiguities, such as tuning effect and fizz-gas confusion, cannot be eliminated without new information. The correct use of seismic attributes also plays a vital role.