Investigation and comparison of conventional methods for estimating shear wave velocity from well logging data in one of the sandstone reservoirs in southern Iran

Estimation of compressional and shear wave velocities is very important in the oil and gas industry. Unlike compressional wave velocity, shear wave velocity is not measured in all wells of a field due to its higher costs. Therefore, using an alternative method that estimates the shear wave velocity at a lower cost and with acceptable accuracy is inevitable. In this study, to estimate the response variable in a well, the correlation of several logs in that well (i.e., acoustic logs, density, neutron porosity, resistivity, gamma ray, dolomite volume, quartz volume, and water saturation) with target log investigated. It was found that the compressional wave velocity, density, dolomite volume, and quartz volume logs are more correlated with shear wave velocity. Therefore, these logs were selected as input features for estimating shear wave velocity using different approaches. In the next step, among the various methods, The estimated values obtained from a method that has the best match with the actual shear wave velocity is introduced as the optimal model. Afterward, it is performed to estimate the shear wave velocity in other wells that do not have a shear wave velocity log. In this paper, multiple regression methods and machine learning algorithms (support vector regression, adaptive Neuro-fuzzy inference system, and deep artificial neural network) were applied to predict the shear wave velocity. In this study, data from seven wells were used. Due to the fact that only in well #7 shear wave velocity has been measured, and in six other wells this feature has not been recorded, this field data limitation has caused the data of well #7 to be divided into training, testing, and validation data. In multiple regression methods (linear and interaction models), support vector regression, and adaptive Neuro-fuzzy inference system, Randomly, 70% of the data has been used for training and 30% for testing, but in the artificial neural network method, Randomly, 70% of the data has been used for training, 15% for validation and 15% for network testing. For all methods, the root means square error and correlation between actual and estimated data are calculated. Linear model, interaction model, support vector regression, adaptive Neuro-fuzzy inference system, and deep artificial neural network have provided 91, 92, 89, 94, and 98% correlation in training data, and 88, 89, 86, 90 and 92% in testing data, respectively. Also, the RMSE for each of the mentioned methods is 125.59, 115.86, 148.23, 84.36, and 80.49 (m/s) in the training data and 139.77, 133.44, 166.03, 126.15, and 98.04 (m/s) in the testing data, respectively. Our results show that deep artificial neural network has provided a better solution than other methods. Hence, in this study deep artificial neural network has been proposed to estimate the shear wave velocity in other blind wells. Moreover, the Castagna empirical model was used to validate the obtained results from the deep artificial neural network in these wells, which show a good fit between the two models.