Velocity model building using transfer learning

Abstract

Deep learning is a promising approach to velocity model building, for it has the potential of processing large seismic surveys with minimal human intervention. By leveraging large quantities of model-gathers pairs, neural networks can automatically map data to the model space, directly providing a solution to the inverse problem. However, such mapping requires big data, which proves prohibitive for 2D and 3D surveys of realistic size. To tackle this problem, we propose a transfer learning strategy: a network is first trained on a smaller subproblem which then becomes the starting solution to a larger, more difficult dataset. We perform transfer learning by first training a neural network at estimating horizontally layered velocity models and then proceed to train an augmented network at estimating 2D dipping layered models. Results show that transfer learning allows using smaller amounts of 2D models and improves convergence. The structural similarity index measure (SSIM) on estimated 2D interval velocity models in time domain has mean 0.8400 and standard deviation 0.0729, or a RMSE of mean 310.1 m/s and standard deviation 110.7 m/s.