Robust seismic inversion workflow using temporal convolutional networks: a case study on noisy field data

Abstract

Summary Seismic inversion, where rock properties are estimated from seismic reflectivity data, typically performs poorly in the presence of noise. The ability of deep neural networks to learn complex relationships offers a potential alternative to conventional inversion. Temporal convolutional networks (TCN), a type of sequence modelling architecture, have previously been applied for seismic inversion on synthetic datasets with promising results. In this paper, we extend this work to test the performance of a TCN on a field dataset contaminated with strong coherent noise. The machine learning approach was found to outperform the conventional inversion result in this case, learning to ignore the noise events. Key to these findings was the development of a realistic synthetic dataset to provide enough samples for training the deep neural network.