A non-subjective technique is presented to combine seismic based fracture detection from shear wave anisotropy with traditional well and structure data to produce a new generation of fracture intensity maps and volumes. Neural networks are easy to use and extremely useful for nonlinear regression where the functional form of the nonlinear equation between the independent and dependent variables are not known. We use a neural network for nonlinearly regressing a number of independent variables to predict fracture intensity within a reservoir. Several of the independent variables are derived from an amplitude versus angle and azimuth (AVAZ) process applied to pre-stack, P-wave seismic data. By examining the changes in reflectivity amplitude with respect to azimuth and incident angle on pre-stack seismic gathers, we can detect the presence of open, near-vertical fractures and their orientation. Seismic data has good lateral coverage, but poor vertical resolution versus well data, which has good vertical resolution, but is sparse laterally. The seismic attributes extracted from the AVAZ process are combined with 3D model attributes, such as porosity and lithology and structural attributes, such as the first and second derivatives of the structural surfaces to predict fracture intensity. Because fracture intensity information is sparse and difficult to obtain, we use expected ultimate recovery (EUR) as a substitute for fracture intensity. The four-month cumulative production provides a good estimate for EUR and is used as the fracture indicator in this study. The attributes are ranked according to their correlation with fracture intensity, and a subset of variables is selected as input into the neural network for the prediction of the fracture indicator. Multiple neural networks are trained on the data providing multiple solutions. These solutions, in the form of 2D maps or 3D volumes, are analyzed statistically to predict the fracture indicator and to high grade prospective drilling locations. In addition, the fracture indicator maps and volumes can be input into discrete fracture modeling tools for further analysis.
Introduction
The purpose of this project was to develop a patented workflow for processing seismic amplitude versus angle and azimuth (AVAZ) results through a neural network to generate maps and volumes of fracture intensity.
The AVAZ seismic processing algorithm analyzes the change in amplitudes with offset and with angle to define fracture orientation and intensity. The neural network (NN) application ranks a number of maps or 3D volumes of reservoir properties with a fracture indicator, allows the user to select which reservoir properties or "drivers" should be used to predict fracture intensity, creates multiple neural networks that relate the "drivers" to fracture intensity, and produces maps or volumes of fracture intensity. In essence, the neural network acts as a non-linear regression relating the "drivers", or independent variables, to fracture intensity, the dependent variable. The advantage of a neural network is that the user does not need to know the form of the nonlinear relationship to be able to fit a function and use it for prediction.
The NN application outputs maps or 3D volumes of fracture intensity depending upon the form of the input data. True fracture intensity is rarely available, but in some tight gas reservoirs, production is related to the amount of fracturing in the reservoir1,2. In these cases, we use a measure of production as a proxy for our fracture indicator. Therefore, the final maps or 3D volumes output out of the NN application will be predictions of production.
One of the problems with using production information as a fracture intensity measure is that other factors can influence production in addition to the natural fractures. For example, operators have been drilling wells on the Pinedale anticline for more than half a century, yet the gas play has not been successfully and sustainably exploited until the past six years due to advances in well completions and frac technology. Since we are not analyzing the completions strategy as part of this study, well completions will be one of the sources of uncertainty in the final results.
