Stable clustering of offshore downhole data using a combined k-means and Gaussian mixture modelling approach

We use unsupervised machine learning techniques aided by the Gaussian mixture model (GMM) for clustering downhole data of gas hydrate reservoir in the Andaman Sea, where drilling and coring were done at Site 17 in 2006 under the first expedition of the Indian National Gas Hydrate Programme (NGHP-01). Six different logging data (namely; density, neutron porosity, gamma-ray, resistivity, P- and S-wave velocity) are used in this study. We obtain six clusters using Davies-Bouldin index, Calinski-Harabasz index, Dunn index, dendrogram and self-organizing map, which are verified by high silhouette values. Data are then clustered using k-means, principal component analysis (PCA) and GMM. We notice that the k-means with random initialization gets biased towards the dominant principal component (gamma-ray), whereas, PCA shows each log has optimal weightage. Based on statistical analysis using 100 runs, GMM with the k-means initialization provides better results than GMM with random initialization. However, it provides three possible configurations of six clusters, which become stable when a combination of six logs is used as another input. Six clusters are interpreted in terms of lithology by histogram analysis of the corresponding log values. Lithology is found clay-dominated sediments with little silt and sand, as well as scattered volcanic ash, carbonate ooze, and pyrite, which is consistent with the lithology determined by smear slide and sieve data. Except at a few depths with higher concentrations (20–50%) in volcanic glass and carbonate ooze, gas hydrate occupies about 10% of the pore space in silty-clay sediments with sand and volcanic ash.