Log1DNet: A Deep Learning Architecture for Sonic Log Prediction for Seal Rock Identification in Carbon Capture and Storage Projects

Abstract

This research presents a one-dimensional Convolutional Neural Network (CNN) architecture for compressional and shear sonic logs prediction to identify potential sealing rock formation for successful carbon capture and storage (CCS) projects. Sonic logs are useful geophysical tools in the geomechanical assessment of rock layers and aid in the delineation of potential confinement and containment formations for CO2 storage in depleted reservoirs. However, these logs are usually missing in old depleted fields due to the cost of acquisition, cycle skipping, or poor borehole condition. Therefore, a deep learning approach is proposed to predict compressional and shear sonic logs, simultaneously. Utilizing open-source data from the decommissioned Volve field in the Norwegian North Sea, Log1DNet, a fully-connected CNN model was employed to capture the trend of sonic log responses. A total dataset of 47,041 is gathered from five wells within the 15/9 block of the field (15/9-F-1A, 15/9-F-1B, 15/9-F-11A, 15/9-F-11T2, and 15/9-F-4). Wells 15/9-F-1A, 15/9-F-1B and 15/9-F-11A were used to train and validate the model while wells 15/9-F-11T2 and 15/9-F-4 served as the test wells, achieving an accuracy of up to 90% when compared with ground truth data. By analyzing various zonation behaviours, the detected zones were leveraged to comprehend the findings of the neural network prediction and delineate zones that can serve as a potential seal for CO2 storage. This approach enables a faster CCS evaluation workflow characterized by low cost and high accuracy, offering significant benefits for the effective implementation of CCS initiatives.