Feasibility of Calculating Petrophysical Properties in Tight Sand Reservoirs Using Neural Networks. Final Report, October 1989-July 1991

摘要

The objective of the research was to determine the feasibility of using neural networks to estimate petrophysical properties in tight sand reservoirs. A second objective was to gain some experience concerning how to approach the development of a future prototype, including what should be done and what should be avoided. Gas Research Institute (GRI) focused the project on tight sands because they contain enormous gas reserves and their complicated lithology represents a challenge to log analysts. The data were supplied by GRI from two of its geographically proximate experimental wells in tight sand formations. The nets were tested in sections of those wells that were not used for training, and in two other wells, one in a geographically close but geologically unrelated formation and one in Wyoming. The feasibility testing demonstrated that the relatively simple neural networks developed have comparable accuracy with standard logging analysis estimates in wells that contributed data to the training set. Transportability of the network was tested by using core measurements in two wells in which the nets were not trained, with inconclusive results. Recommendations were made to increase the accuracy of the neural networks.