Subsurface model prediction using a neural network - A real data example from the Rock-Springs uplift, Wyoming

Seismic inversion is a method which extracts seismic attributes such as the P- and S-wave velocities, density, acoustic impedance, Poison's ratio etc. from seismic data, which in turn, provide the subsurface rock physics i.e., the reservoir attributes such as porosity, permeability, lithology etc. Conventional model based inversions, such as Prestack waveform inversion (PWI), simultaneous amplitude-variation-with-offset (AVO) inversion, use theoretical relationships between model parameters and observed seismic data that are based upon some assumptions. All these methods predict the subsurface earth model to an accuracy to which the underlying theoretical relationships are valid. In this paper we present a new approach of seismic inversion methodology, which is a data driven methodology (Schultz et al., 1994), and uses statistical approach rather than deterministic method to predict the subsurface reservoir properties. This method combines neural net inversion with PWI in a hybrid methodology. Applying our method to the Rock-Springs uplift (RSU) real seismic data and comparing with AVO inversion, we demonstrate that our method provides a comparable image to that from AVO inversion.