Shear wave velocity prediction using Elman artificial neural network

Abstract

Shear wave velocity (Vs) is one of the most important features in seismic exploration, reservoir development and characterization. Conventionally, Vs is obtained from core analysis which is an expensive and time-consuming process, and dipole sonic imager (DSI) tools are not available in all wells. Therefore, developing a fast, cheap, and reliable alternative way to Vs prediction with continuous values makes a great contribution in reservoir characterization. In this study, a new method as an alternative way is proposed based on Elman artificial neural network to predict Vs using well log data including gamma ray (GR), resistivity (LLD), neutron porosity (NPHI), bulk density (RHOB), compression wave velocity (Vp), and water saturation (Sw). Elman network is a memorized network and has a feedback from each hidden layer to the former one, and the subsequent behavior can be shaped by the previous responses. Levenberg–Marquardt training algorithm is used to optimize the weight and bias values, and fivefold cross-validation is also used to ensure that the developed ANN is not prone to over fitting. In order to compare the ability of the proposed method with the other common methods, three empirical relations including Castagna, Brocher, and Carroll, and also three other ANNs with such topology including MLP and Elman and such training algorithms as particle swarm optimization are used. Eventually, based on the training time, and based on the best measured mean squared error (MSE) and R2 for our test data, the proposed Elman network is more accurate, efficient, reliable, and robust than the other three empirical relations and three ANNs. Our experimental results demonstrate a successful adapting for Vs prediction using Elman ANN. The proposed method could be applied in important applications such as geomechanical and AVO modeling.