Predicting lithology using neural networks from downhole data of a gas hydrate reservoir in the Krishna–Godavari basin, eastern Indian offshore

The Bayesian Neural Network along with hybrid Monte Carlo search technique (BNN-HMC) plus multivariate data analysis (K-means, dendrogram, 3-D cluster data analysis, PCA, SOM) is successfully employed to well log data at three holes for the classification of rock-type/litho-type successions in the KG basin, eastern Indian offshore. Seven unsupervised classification techniques successfully determine the number of lithologies and also detect the presence of gas hydrate as a unique cluster at all three holes. The technique maps various litho-units and gas hydrate with depth with 99% correlation between output and the target. Our results reveal that the study area is dominated by clay and silt with minor amount of silty clay and sand. The combined approach of unsupervised and supervised classification techniques is found to be very effective to detect and delineate gas hydrate distribution. Gas hydrate is found to be distributed mainly in clay, silty clay and silt, not in sand. Our results clearly illustrate that if gas hydrate is not considered as a separate unit, it will be distributed as a lithology in its hosts and identification of lithology will be erroneous. The distribution of permeability along with gas hydrate saturation and litho-section helps to recognize the permeable and impermeable layers present at three holes in the KG basin. The estimated permeability is very low (10−22 to 10−19 m2) mainly due to the presence of clay. This approach is found to be very useful than any other techniques for the identification of lithology in a gas hydrate reservoir. 

Figure: Correlation of predicted lithology with seismic litho-facies



Amrita Singh, Maheswar Ojha and Kalachand Sain; Geophys. J. Int. (2020) 220, 1813–1837 https://doi: 10.1093/gji/ggz522 GJI Marine geosciences and applied geophysics