Estimation Of Reservoir Properties From Seismic Data By Smooth Neural Networks

Abstract

Traditional joint inversion methods reqnire an a priori prescribed operator that links the reservoir properties to the observed seismic response. The methods also rely on a linearized approach to the solution that makes them heavily dependent on the selection of the starting model. Neural networks provide a useful alternative that is inherently nonlinear and completely data-driven, but the performance of traditional back-propagation networks in production settings has been inconsistent due to the extensive parameter tweaking needed to achieve satisfactory results and to avoid overfitting the data. In addition, the accuracy of these traditional networks is sensitive to network parameters, such as the network size and training length. We present an approach to estimate the point-values of the reservoir rock properties (such as porosity) from seismic and well log data through the use of regularized back propagation and radial basis networks. Both types of networks have inherent smoothness characteristics that alleviate the nonmonotonous generalization problem associated with traditional networks and help to avert overfitting the data. The approach we present therefore avoids the drawbacks of both the joint inversion methods and traditional back-propagation networks. Specifically, it is inherently nonlinear, requires no a priori operator or initial model, and is not prone to overfitting problems, thus requiring no extensive parameter experimentation.