Machine learning in the classification of lithology using downhole NMR data of the NGHP-02 expedition in the Krishna-Godavari offshore Basin, India

We apply machine learning techniques to classify the lithology using downhole nuclear magnetic resonance (NMR) data of gas hydrate reservoirs in the Krishna-Godavari (KG) Basin, eastern Indian offshore. We choose two sites of the second expedition of the Indian National Gas Hydrate Program (NGHP-02) for our study, one with clay dominated site NGHP-02-01 from Area E and another with silt/sand dominated site NGHP-02-05 from Area C. We obtain five distinct classes in NMR signals at two sites using unsupervised techniques such as Davies-Bouldin index, silhouette, self-organizing map, k-means clustering and emergent self-organizing map. Obtained classes are interpreted using the NMR T-2 distribution pattern with time in terms of clay, silt, sand and pebble/gravels with depth and reconciled with gamma-ray, density, porosity, resistivity, velocity and photoelectric factor. Our results demonstrate the lithology as silty-clay with minor sand at Hole NGHP-02-01A, and silt, sand with less clay and minor pebble/gravel at Hole NGHP-02-05A. The presence of gas hydrate is identified by observing an overall low amplitude T-2 distribution curve, which is due to the reduction of pore space by solid gas hydrate. The maximum concentration of gas hydrate of about 43% of pore space estimated using NMR- and density-porosity are distributed mainly in silty-clay at Hole 01A and of about 75% in clayey-silt and silty-sand at Hole 05A. Low amplitude NMR T-2 signals observed in clay-dominated sediments at higher relaxation time may be due to the fracture-filled gas hydrates at both holes. Interestingly, high porosity (90-95%), high permeability (4 mD), low bulk-density (1.2 gm/cm(3)) and low photoelectric factor (0.8 b/electron) are found in sandy layers within the gas hydrate-stability zone at Hole 05A, which possibly comprising buried channels. Comparison of the lithology obtained from other log data with that from NMR data and core samples illustrates that the NMR data are most effective for characterization of reservoirs, except for a high concentration of gas hydrate in a sandy/coarse-grained reservoir, which needs further studies.

Automated summary

This study applies machine learning techniques to classify the lithology of gas hydrate reservoirs in the Krishna-Godavari Basin, India. By analyzing NMR data and interpreting the obtained classes, the distribution of different lithologies and the presence of gas hydrates are determined.