Evaluation of Radial Basis Function Neural Networks in Reservoir Characterization of Caddo Member in Boonsville Field, Texas
Tao Zhao and Kumar Ramachandran
Geophysical reservoir characterization requires building a nonlinear relation between seismic attributes and rock/fluid properties computed from well logs. With such a relation, the rock/fluid properties computed from well logs can be extended to inter-well points. Neural networks can be employed to obtain this nonlinear relation. In this study, radial basis function (RBF) neural networks are evaluated in the application of porosity prediction. The structure of a typical RBF network is composed of an input layer, an output layer and a hidden layer. Currently, RBF network is only used as single hidden layer network in geophysics applications; however, multilayer RBF networks have already been dealt with by some researchers (Chao et al., 2001) and according to their study, there exists a performance improvement when multiple hidden layers are used. This study explores the possibility of applying multilayer RBF networks in reservoir characterization, dealing with well logs and seismic data and to design an optimal structure for a RBF network with fixed number of nodes. The seismic and well log data used in this study are the public part of the Boonsville 3-D seismic dataset, which is from the Boonsville field in north central Texas. A series of tests are carried out to examine performance and inspired by the comparative results of RBF and multilayer perceptron (MLP) networks, a hybrid of RBF and MLP called centroid based multilayer perception (CMLP) network is employed for porosity prediction. Finally, the best CMLP network is used for porosity prediction. Porosity distribution map constructed from seven seismic attributes using a triple layer CMLP neural network shows good correlation with well data. Because of the assumptions and approximations during the processes of porosity log prediction, porosity downscaling and neural network prediction, the average porosity prediction error is around 20%.
AAPG Search and Discovery Article #90176©AAPG Mid-Continent Meeting, Wichita, Kansas, October 12-15, 2013