Amplitude-variation-with-offset, prestack waveform, and neural network inversion- A comparative study using real data example from the Rock Springs Uplift, Wyoming

In this work, I use seismic and well data to predict the subsurface geologic model. To accomplish this task, three approaches have been used: (1) amplitude-variation-with-offset inversion, (2) prestack waveform inversion, and (3) neural net inversion. Both amplitude-variation-with-offset and prestack waveform inversion are model-based inversions in which an initial (guess) model is iteratively modified until the synthetic (predicted) data from the model and underlying physics matches observation to reasonable accuracy. While the amplitude-variation-with-offset inversion uses a convolutional model for the underlying physics for computing synthetic data, the prestack waveform inversion uses a rigorous wave equation-based method to compute them. The neural network inversion, on the other hand, is a data-driven inversion methodology in which the network undergoes a series of training to derive linear/nonlinear relationships between the seismic attributes and the model attributes. Once such relations are established, they are used to predict the subsurface model directly from the seismic data. Using all three inversion methods on a single dataset from the Rock-springs uplift, Wyoming, it has been found that both amplitude-variation-with-offset and neural net inversions produce comparable results although the subsurface model estimated by the latter is of slightly higher resolution than the former. Prestack waveform inversion, even though compute-intense, is far superior to the other two inversion methods and should be the method of choice as the parallel computers with large number of compute-nodes become commonly available.