Seismic data conditioning and neural network-based attribute selection for enhanced fault detection

Abstract

In this study of the Dorood oil field, offshore Iran, 3D seismic data were utilized to identify a complex fault pattern in the highly faulted and fractured Fahliyan Formation. To enhance data quality and improve attribute accuracy and detection power, a steering cube was first computed based on a sliding 3D Fourier analysis technique, using the concept of directivity. The steering cube, which contains dip and azimuth information for each trace, was utilized for calculation of dip-steered filters and attributes. We applied the dip-steered median filter to remove random noise and to enhance laterally continuous seismic events by filtering along the structural dip. Several fault identification attributes, such as dip, curvature, coherency and similarity, and a meta-attribute of a ridge enhancement filter, were extracted from dip-steered, noise-attenuated data. A supervised, fully connected multi-layer perceptron neural network was constructed to select and combine the most sensitive fault attributes. The neural network, trained at identified fault and non-fault locations, was applied to the whole seismic volume to generate a cube of fault probability. Interpretations of major faults and fractures were integrated with geological, reservoir engineering and production data to highlight the role of these heterogeneities on dynamic reservoir properties. Faults and fractures in the Fahliyan reservoir were identified in general to have the effect of decreasing reservoir permeability. While most of the faults recognized have locally a sealing capacity effect, when the fault throws are large enough to bring into contact the Manifa and Yamama reservoirs, they act as a conduit for fluid communication and, wherever a well crosses a major fault, early water breakthrough is observable.