Identification of Lithology using machine learning of gas hydrate reservoir in Krishna-Godavari offshore basin, India

Abstract

Summary The presence of gas hydrate increases velocity responses of the well log data, depending on saturation. This may lead to error in identifying lithology using the neural network without understanding the effect of gas hydrate. Here, we have identified the host litho-units of gas hydrate and its effects on identification of lithology by employing the unsupervised and supervised neural network learning techniques together to well log data in the Krishna-Godavari (KG) basin, eastern Indian margin. We have used density, neutron porosity, impedance and Poisson ratio of hole 10A, which was acquired in 2006 under the Expedition-01 of Indian National Gas Hydrate Program (NGHP-01). At first, the unsupervised method (e.g. Davies-Bouldin Index, self-organising map and k-means clustering) has been applied to the data with gas hydrate and without gas hydrate (i.e. water saturated) to obtain the optimum number of clusters/litho-units. The optimum number of clusters having the lowest Davise-Bouldin index value has been taken as input in k-means clustering. Then we have applied the supervised learning technique (Bayesian neural network aided with Hybrid Monte Carlo searching technique) to refine each defined cluster unit of training samples and map them with depth. The gas hydrate zone is recognized as a unique cluster unit with other cluster units present in data and mapped it with depth. The present technique identifies four litho-units dominated by clay with minor amount silty-clay, silt and sand, and gas hydrate is found mainly in clay, silty-clay and silt, not in sand. The results also show that gas hydrate is distributed as a lithology in its hosts litho-units (i.e. clay, silty-clay and silt), if we do not consider gas hydrate as a separate unit. The identified lithologies correlates very well with the seismic section passing through the hole 10A in the KG basin.