An algorithm to predict shear wave velocity using well log data and deep learning algorithms

Shear wave velocity (Vs) is one of the most critical parameters for determining geomechanical properties to predict reservoir behavior. Determining shear wave velocity (Vs) through methods, such as core analysis requires a significant amount of time and cost. Additionally, due to the scarcity of core samples and the heterogeneity of reservoir rocks, determining this parameter using conventional methods is often not very accurate. While many empirical methods have been developed for estimating Vs, their applicability across different regions is often limited. Therefore, estimating Vs using conventional well logs is crucial. An efficient method for predicting Vs is the use of intelligent algorithms, which offer low-cost and accurate predictions. It is feasible to predict Vs using well log data. In this study, Vs was predicted using empirical relations and some deep learning (DL) algorithms in one of the hydrocarbon fields in southern Iran. In order to use the DL methods, the autoencoders deep network was used to select the effective features in predicting the Vs, and then, with multi-layer perceptron (MLP), long-short term memory (LSTM), convolutional neural network (CNN), and convolutional neural network + long-short term memory (CNN+LSTM) networks, Vs was predicted. The performance of these models was tested by a blind data set that the models had not seen before. Furthermore, the results were checked and evaluated by set of statistical measures, including MAE, MAPE, MSE, RMSE, NRMSE, and R۲ values calculated for train, test, and blind datasets. It was found that all four deep learning models used in this study well performed effectively for Vs prediction, but the combined CNN+LSTM model results indicated that the least root mean squared error (RMSE) was equal to ۰.۰۲۴۳ (۲.۴۳%) and the best coefficient of determination (R۲) equal to ۰.۹۹۹۳ for blind dataset. We found that Vs can be predicted from a series of well log data by considering their variation trends and context information with depth by means of DL algorithms. This approach is particularly suitable for problems involving various series data, such as Vs prediction. By comparing the results obtained from DL algorithms with those from conventional empirical methods and processing real petrophysical well log data, it can be concluded that deep learning algorithms not only offer more predictive accuracy and robustness but also hold promising use prospects in Vs prediction studies. The results showed that the used CNN and CNN+LSTM networks, as new deep learning algorithms, are able to predict Vs adequately.

کلیدواژه ها:

Shear wave velocity ، Log data ، empirical relations ، Deep learning algorithm