Estimation brittleness index in carbonate environments using log and lithology data and deep learning techniques

Abstract

Brittleness index is one of the most important Geomechanical parameters, which has a great impact on the rock breaking process and drilling activities. The methods of evaluating brittleness of rocks are mainly divided into three categories: (1) direct laboratory, (2) mineral content, and (3) based on elastic moduli. one of the efficient methods for brittleness index predicting is use of intelligent methods, which are low-cost and accurate methods, and it is possible to predict the brittleness index using log and lithology data. In this study, dynamic and static brittleness index values are predicted using deep learning (DL) algorithms and lithology data in carbonate environment in one of the hydrocarbon fields in southern part of Iran. In this paper, the effective features were selected using the deep learning algorithm of the Auto-encoder, and the dynamic and static brittleness index was estimated using the MLP, LSM, and CNN algorithms. As 12 laboratory core samples were available, at first the brittleness index were calculated by relevant empirical relations and data of some available well logs in order to generalize these core results to the entire target depth range of 3551.07 to 3799.78 meters. Then a set of relationship between the well’s logs derived dynamic and static brittleness index and laboratory results was determined for the depths where the laboratory samples were recorded. Following that, an Auto-encoders deep network was used to select the effective features in predicting the brittleness index, and finally by using MLP, LSTM and CNN networks the value of dynamic and static brittleness index was predicted. Here, the goal is to obtain the brittleness index values with high accuracy wherein there no core data. The performance of the three algorithms prediction models is tested by blind data sets that the models have not seen before. Furthermore, the results were checked and evaluated by set of statistical measures like MAE, MAPE, MSE, RMSE, NRMSE and R2 values that calculated for train, test and blind dataset. At first, dynamic brittleness index estimate using log data and three algorithms and R2 for blind data equal to R2MLP=0.91, R2LSTM=0.97, R2CNN=0.98, in the following, using MLP, LSTM and CNN the dynamic brittleness index has been converted into a static brittleness index and R2 for blind data equal to R2MLP=0.94, R2LSTM=0.96, R2CNN=0.96. Finally, the static brittleness index has been estimated directly from the log data without the relation of dynamic to static transformation and R2 for blind data equal to R2MLP=0.95, R2LSTM=0.96, R2CNN=0.97. Finally, the dynamic and static brittleness index was compared with the brittleness index obtained from lithology, and there is a good match between them. The results show that the deep learning algorithm is a novel method, robustness and accurate method in estimating the dynamic and static brittleness index using conventional logs. The results show used CNN and LSTM networks as new deep learning algorithms to predict brittleness index.