A comparative study of transfer learning methodologies and causality for seismic inversion with temporal convolutional networks

Machine learning-based seismic inversion methods suffer from high labeled data requirement in the absence of which they may fail to generalize. Recent data-driven inversion methodologies based on temporal convolutional networks use 2-D sequence models and transfer learning to unburden the algorithm from high training data requirements. Such methods are restricted in that they only model seismic and well log data causally, contrary to the physics of the inversion process. Moreover, they require all data involved in the study to be of exactly the same sampling and resolution factors, a scenario unlikely to happen in practice. We show that it is possible to extend the method to unequal resolution and sampling factors. Further, we demonstrate that by removing the constraint of causality, we are able to improve the performance of the algorithm. Moreover, we perform a comparative study of various transfer learning methodologies in the literature in the context of data-driven inversion. Using the best performing transfer learning methodology in combination with non-causal networks, we achieve the lowest MSE on the SEAM dataset of 0.0603.