Application of artificial intelligence to eletrofacies identification neural networks versus discriminant analysis

Abstract

Electro-facies are identified by a neural network (NN) trained with well-log and core data. Unlike traditional expert systems, it is unnecessary to feed rufes to an NN, learning being achieved ínstead through changes that occur in the network's connection weights following stímuli input. Furthermore, NNs are well suited for analyzing missing or incomplete data, and they "degrade gracefully" when a portion of the network is destroyed. ln the proposed backpropagation network, gamma-ray, neutron porosity, and bufk density logs, coupfed with core data, are used as input. The errar is propagated backwards from the output fayer to the input fayer, and the connection weights are modified in proportion to the contribution of the connections to the overall error. Once trained, this NN performs as well as discriminam analysis in identifying electric-log facies. The authors belíeve that the combination of neurocomputing and traditional computing methods fike discriminant analysis can help in the solution of many problems in electro-facies identification.