Porosity and Permeability Prediction in Low-Permeability Gas Reservoirs From Well Logs Using Neura Networks

Abstract Artificial neural networks are gaining popularity as tools for estimating reservoir parameters from limited, common data suites. Requirements for their use include input data such as well logs that relate to the desired output, and "truth" data for training. Two case studies will be used to illustrate the use of neural networks to predict porosity and permeability from log data. In both cases, the predictions were needed for field studies aimed at improving reservoir management and optimizing production. In the giant Hugoton gas field in Kansas, porosity was predicted from spectral gamma ray, photoelectric, and bulk density data with generic neural network software. Such software allowed the relationship developed by the neural network to be translated into an equation that could be readily applied to all wells with the requisite log curves. Permeability was predicted using a more log-oriented type of software, one that incorporated depth windows of input data in generating and applying the network. In addition to the input curves used for porosity, neutron logs were used in the permeability prediction. In the Hugoton example, mineralogy was a critical factor in porosity and permeability determination, so most of the input data provided information about mineral constituents of the reservoir. Core analyses served as "truth" cases in training both porosity and permeability neural networks. In the Red Oak gas field in Oklahoma, where density logs are commonly absent or of poor quality, multiple neural networks were developed to predict density from gamma ray and deep induction data. A combination of measured and predicted density curves were then used to calculate porosity. One additional network was built to estimate permeability from gamma ray, induction, and density data over 5 orders of magnitude of permeability, with core plug permeability measurements used as truth in training. Key features of these networks were selective rather than statistical training, back prediction of input data for validation, and the use of vertical intervals of input data. One additional network was built to estimate permeability form gamma ray, induction, and density data over 5 orders of magnitude of permeability, with core plug permeability measurements used as truth in training. Key features of these networks were selective rather that statistical training, back prediction of input data for validation, and the use of vertical intervals of input data. P. 563