Deep neural networks for 1d impedance inversion

Abstract

We investigate the applicability of different types of deep neural networks for the estimation of subsurface properties from seismic data. The pre-trained networks can predict velocity models from new data in a few milliseconds, which makes this data-driven approach especially important for multidimensional inversion, where conventional methods inversion methods suffer from large computational cost. At the same time, realistic one-dimensional models such as the 160-layer velocity model used as an example in this study require large synthetic datasets for training, which are not always possible to obtain. Hence, we also study the impact of extending the training data by adding random noise to the modelled examples. We observe that enlarging training datasets by adding synthetic noise to existing samples improves the quality of inversion without a significant increase in computational complexity.