Abstract
Feedforward neural networks are used to estimate reservoir properties. The neural networks are trained with known reservoir properties from well log data and seismic waveforms at well locations. The trained neural networks are then applied to the whole seismic survey to generate a map of the predicted reservoir property. Both theoretical analysis and testing with synthetic models show that the neural networks are adaptive to coherent noise and that random noise in the training samples may increase the robustness of the trained neural networks. This approach was applied to a mature oil field to explore for Devonian reef-edge oil by estimating the product of porosity and net pay thickness in northern Alberta, Canada. The resulting prediction map was used to select new well locations and design horizontal well trajectories. Four wells were drilled based on the prediction; all were successful. This increased production of the oil field by about 20%.