Successful recovery of hydrocarbons from either newly discovered or existing fields depends upon the completeness and accuracy of the reservoir characterization. Wellbore logging, the fundamental characterization technique currently in use, can in theory provide sufficient data to determine the principal parameters needed to evaluate a reservoir. In application, well logging is most successful in the homogeneous reservoirs where definitive signatures are embedded within log suites that allow positive identification of hydrocarbon-bearing zones in a wellbore. In heterogeneous reservoirs, hydrocarbon signatures in log suites are difficult to identify because of the variability of the signal response in complex lithologies. This projeces Phase I effort investigates the utility of applying neural networks to recognize and exploit signatures embedded in a suite of wellbore logs for wells in heterogeneous reservoirs. Essentially, the networks are trained to predict which zones in a well will be productive, to estimate the potential productivity of those zones, and to estimate key reservoir parameters such as porosity and permeability. The data used for Phase I are obtained from the complex Silurian Interlake formation of the Williston Basin in North Dakota and consist of the existing conventional well logs available. Key questions being answered during Phase I include how accurately the networks predict the presence of hydrocarbons in bore holes, how much wellbore information is required for accurate models, and how the methodology can be extended to predict parameters such as porosity and permeability. Additionally, Phase I includes examination of how data from other sources, i.e., drill-stem tests, core analysis, and mud logs, can be integrated into the pattern recognition methodology, and to what extent such data improve the network's ability to characterize heterogeneous reservoirs. The principal effort in Phase I is directed at identifying which neural networks are applicable to the problem, developing a process by which to train these networks, and ultimately demonstrating the viability of utilizing a neural network-based concept to enhance reservoir characterization.Anticipated Results/Potential Commercial Applications as described by the awardee:Phase I will demonstrate the utility of a neural networkbased pattern recognition methodology that is empirical, independent of heuristics, and does not require expensive re-entry of the borehole to collect additional information. Phase I also examines how the methodology can be extended to predict reservoir parameters such as permeability. The experience and systems developed in Phase I are to be extended and tested during Phase II on other heterogeneous reservoirs to confirm conclusively that the process is broadly applicable. Phase II results will be a demonstrated methodology that can be marketed as a service and potentially bundled as a hardware/software solution.
