Robust deep learning-based seismic inversion workflow using temporal convolutional networks

Seismic inversion is the process of converting seismic reflectivity data into physical subsurface properties. The most common inversion methods use physics-based forward modeling, but these require time-consuming steps, such as initial model building and wavelet extraction. Coherent noise in the seismic volume also may lead to suboptimal results. Advances in deep learning enable the development of new geophysical workflows that may help overcome these challenges. One example is the temporal convolutional network (TCN), a deep neural network that learns from sequential data, such as seismic traces. Previous research using the TCN architecture has indicated promising inversion results on synthetic data. However, applying the method to field data has several additional challenges that need to be considered, including complex noise and limited well availability. We used a poststack field data set containing coherent noise to evaluate the TCN approach for acoustic impedance inversion under these conditions. Despite the small data set, a TCN trained using traces and logs acquired at well locations produced better results than conventional inversion when supplemented with an additional time feature. While the physics-based inversion created false artifacts related to the noise, the neural network approach learned to ignore the suspected multiple events. Supervised learning using well data also makes semi-automated inversion a possibility. However, obtaining acceptable results using the few locations with logged boreholes may only be possible in relatively simple geological scenarios. To overcome the issue of small data sets, we developed a workflow for generating realistic synthetic data to provide more samples and variation for model training. A TCN trained using synthetic data ultimately produced the best impedance estimates, but care is needed to ensure the synthetic traces contain realistic noise. Overall, we show that a TCN can successfully invert seismic data contaminated with coherent noise, producing superior results compared to model-based inversion.