Abstract

Lithofacies classification is a critical component of geological investigations, traditionally performed manually by experienced geoscientists using well logs, core data, and master logs. Accurate classification and understanding of facies distribution in 3D space provide valuable insights into depositional environments and rock properties such as porosity, permeability, and water saturation, ultimately aiding in better reservoir characterization. This study presents a semi-automated lithofacies classification workflow using a machine learning approach, aimed at reducing manual interpretation time and minimizing subjective biases. The study was conducted in an oil field of Cambay Basin using an open-source data mining tool. The dataset includes 53 wells, comprising 63,273 labelled observations with associated measured properties at various depths. Initial analysis involved ditch sample and well log interpretations to estimate the potential number of facies. An unsupervised classification technique was initially applied using data from 53 wells to explore the maximum number of facies present in the study area, based on selected input log curves. The resulting clusters from the unsupervised clustering process were interpreted as preliminary facies classes. These initial classifications were subsequently validated and refined through rigorous quality control by domain experts, incorporating geological insights to ensure a consistent and geologically robust facies interpretation. This was followed by supervised machine learning for facies classification and prediction in remaining blind wells.