-
oa Estimation of Reservoir Properties from Seismic Attributes and Well Log Data Using Artificial Neural Networks
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, GEO 2010, Mar 2010, cp-248-00074
Abstract
Porosity, permeability are key factors to build a 3D geological model for a reservoir. The best method<br>to get these properties would be to measure them on core samples in the laboratory. However, this<br>method is costly and time consuming, and usually only a few out of all wells are cored and even then<br>only a small portion of the well. To fill the gap in the vertical scale, geologists generally use a statistical<br>approach, such as linear or non-linear multiple regressions to correlate reservoir properties with the<br>continuously recorded well log data. Recently, geoscientists have utilized Artificial Intelligence (AI),<br>especially Neural Networks (ANNs), to predict reservoir properties. This talk reports a comparative<br>study of two types of neural networks, a Multiple-Layer Perception MLP, with back propagation neural<br>network, and a General Regression Neural Network GRNN. The viability of these techniques are<br>demonstrated on well log data and seismic attributes from sand stone reservoir in south of Algeria.<br>This study utilizes the basic logs including gamma ray GR, interval transit time DT, shale volume VSH,<br>bulk density RHOB, deep later log LLD and corrected porosity NPHI and five attributes( instantaneous<br>frequency, instantaneous phase, RMS amplitude, half energy and Arc length) to predict porosity,<br>permeability and lithofacies in cored and uncored wells. The agreement between the core data and the<br>predicted values by neural networks demonstrate a successful implementation and validation of the<br>network’s ability to map a complex non-linear relationship between well logs and permeability and<br>porosity. Also the results show that the application of the General Regression Neural Network GRNN<br>gives a relatively better performance than the Multiple-Layer Perception MLP.