ETD

• neural network
• seismic inversion

Pre-stack seismic inversion in one of the most used methods to obtain information about the subsurface from seismic data. In the context of petrophysical inversion, pre-stack seismic inversion methods often as a first step, derive the elastic properties(i.e. P-, S- wave velocities and density) and then proceed to obtain the petrophysical properties of interest with a properly calibrated rock physics model. In this thesis, we explore the possibility of utilizing an alternative approach to the petrophysical inversion problem with a well-known class of recurrent neural networks, the Long-Short Term Neural Network. More specifically the task is solved using a variant of the Long-Short Term Memory (LSTM) Neural network known as the Bi-directional LSTM (Bi-LSTM) neural network. Bi-LSTMs are known to improve model performance, and they are well suited for problems where timesteps of the input sequence are available. For this reason, a Bi-LSTM network is trained and subjected to several robustness tests with aim of assessing the performance and suitability of the neural network approach compared to a standard approach i.e. the Linearized Gauss Newton AVA inversion. The Monte Carlo error propagation scheme is also employed to examine how uncertainty in the data space and model approximations propagate unto the predictions derived from both methods.