Generation of missing well log data with deep learning: CNN-Bi-LSTM approach

Well log data is generally collected by the drilling process, which is associated with huge costs and is also time taking. Furthermore, distorted data are widespread in well logs due to instrument damage, poor borehole conditions, imperfect logging, and so on, causing data loss leading to poor interpretation. The missing well log data can be retrieved using deep learning methods from the existing/ available borehole logs. In this study, we propose a Convolutional Bidirectional Long short-term memory (CNN-Bi-LSTM) with fully connected layers that could successfully predict the missing log data for two sites in the Krishna Godavari basin, namely, NGHP-01-14 and NGHP-01-06. In NGHP-01-14, the CNN-Bi-LSTM was employed to predict the S-wave log using the density and gamma logs from the same NGHP-01-14 site. Whereas, in NGHP-01-06, the sonic log is predicted using different logs from the nearby NGHP-01 sites. This method reliably extracts the important features in the logs along the depth of the borehole, which helps to predict the missing data and also the logs that are not available in the well. The accuracy of the predicted data is calculated with an error metric, and the log predicted using CNN, Bi-LSTM, and ANN network results are compared to establish the efficacy of the proposed method. The MSE value of the predicted shear wave log of NGHP-01-14 from the proposed network is 0.0025, and from CNN, Bi-LSTM and ANN are 0.003, 0.0045 and 0.0084, respectively. The error values of the predicted sonic log of NGHP-01-06 from CNN-Bi-LSTM, CNN, Bi-LSTM, and ANN are 0.0025, 0.004, 0.005, and 0.0065, respectively. The outcomes from the network establish that the proposed method predicts the missing log successfully and efficiently.