Assessment of borehole Nuclear Magnetic Resonance data of gas hydrate bearing sediments using neural network technique in eastern Indian offshore

Abstract:

It is obvious challenging to analyse, predict and model voluminous geophysical data accurately in one frame with an efficient time limit. An automated technique, called Machine learning is found to be more effective to handle large as well as complex geophysical data. Here, we have analyse T2 distribution curves of nuclear magnetic resonance (NMR) downhole data with machine learning techniques for mapping of lithology of a gas hydrate reservoir at site NGHP-02-05A in Krishna-Godavari offshore basin, India. At first, we subdivide NMR curves into optimum number of classes using self-organising map (SOM) learning technique to map the different classes. NMR curves represent specific pore size distribution of the rock formation, which could be interpreted in terms of lithology with its relative amount for characterization of gas hydrate reservoir using machine learning techniques. We apply Devies-Bouldin index, SOM and K-means clustering to subdivide NMR curves into optimum number of classes with a similar and unique signal shape and interpreted them in terms of lithology. For classification, we assume clay, silt and sand, as these are the main three litho-units of sedimentary rocks in general. Our results show five types of classes along with gas hydrate bearing layers, which are identified by low signal amplitude of all relaxation times at Hole 05A. Most importantly, subtypes of gas hydrate bearing sediment with its host litho-units are also easily identified. The method, we use here is an automated way to examine NMR curves for classification of sediment types.