Petrophysical Inversion of seismic dataset using artificial neural networks

One of the primary goals of active source seismic surveys is to estimate petrophysical parameters that can aid in the evaluation of reservoir attributes. This is generally done in a two-step process. Solving the non-linear seismic inversion problem to obtain the subsurface velocity profile is the first step followed by including apriori rock physics problem to constrain the petrophysical parameters. Due to the problem's non-unicity, both the procedures necessitate rigorous data analysis. We have attempted to solve the above-mentioned two-step problem in a single step by leveraging the function approximation property of the neural networks. Our approach involved the petrophysical inversion of seismic trace data to obtain the subsurface volume fraction of shale, water saturation, and porosity distribution. The neural network was trained using the synthetic dataset generated by considering the presence and the effect of water- and oil- saturations on layers of sandstone and shale, that affect the elasticity moduli and the bulk density of each layer and hence, can be used to obtain the p-wave velocity and eventually the acoustic impedance. This allowed the inclusion of apriori information from the rock physics models to generate the seismic data, which can then be used to predict petrophysical properties directly from the seismic traces. By following this approach of inverting seismic data to directly obtain the petrophysical parameters, we combine both the steps into one. The earth models considered while generating the training set were kept as realistic as possible by including different sandy shale layers along with the water-bearing sand zone, a cap rock stratum and the most important oil-bearing sand zone. The efficacy of our neural network model is demonstrated in delineating these important zones while also retrieving the petrophysical parameters for all the layers accurately.