Seismic attribute calibration using neural networks

Abstract

Neural networks are used to predict sand percent from seismic attributes for several intervals in a Cretaceous basin. Neural networks are highly simplified computer models of biological neural systems and have found applications in a number of areas including pattern recognition, classification, and signal processing. These networks are not programmed but rather are trained by repeated presentation of input data (seismic attribute sex tracted at well locations) and the corresponding desired output (sand percent measured in the wells). In this context of seismic attribute analysis, training a network is equivalent to a calibration. Nineteen seismic attributes related to reflection continuity and geometry, amplitude, and frequency are extracted from the seismic data. The neural network calibration of these attributes to sand percent derived at 11 well locations shows that, in general, this rock property can be estimated away from well control to within the well measurement accuracy using seismic data. Such a neural network approach should be considered for seismic attribute calibration if there are a large number of attributes to analyze and where the correlations between individual attributes and the desired rock properties is weak.