Permeability characterization and spatial modeling in complex reservoirs: use of tree classifiers and Markov Random Field

Abstract

This research presents two approaches for working with reservoir properties. The first is the application of decision tree classifiers for predicting partitioning or classifications based on well logs for improving the permeability estimations. The second approach is the application of Markov Random Field (MRF) for determination of a 3D high-resolution model of porosity using well logs and seismic attributes. The prediction of permeability is a critical aspect of reservoir characterization in complex reservoirs such as carbonate reservoirs. In order to improve the permeability estimation in these reservoirs, several techniques have been proposed to partition well log response into different classes, viz., layering or zoning, electrofacies and hydraulic flow units. It has been shown that the partition approaches are quite promising. However, they are generally defined using a specific set of well logs. For this reason, difficulties arise in the estimation of permeability when there are missing well logs. In this research we developed an approach, using the decision tree classifiers, for solving the problem of missing well logs. We predicted the partitions and estimated the impact of missing well logs at two complex carbonate reservoirs in west Texas, viz., Salt Creek Field Unit (SCFU) and North Robertson Unit (NRU). Porosity is another important reservoir property that is obtained from different sources in an oil field; the most common source being well logs. In some reservoirs porosity data from well logs is available only for few wells. Reservoir porosity models built using sparse well information can be highly uncertain. Therefore, we need to integrate information from different scales or sources to get an accurate spatial distribution for porosity. A common source of such additional information is seismic data. There are several techniques available for integrating seismic and well log data e.g., cokriging and its variants. These techniques are reasonably accurate but only when there are enough wells for determining the semivariogram from well logs. In this research, the concept of Markov Random Field (MRF) was used to solve the problem of reservoir not having enough wells for applying conventional geostatistical techniques of integration. We generated a 3D high-resolution model for porosity using neutron logs and the maximum seismic amplitude for a reservoir in the Middle East.