Summary We investigate deep learning approaches to inversion of a 1D model of the subsurface using synthetic surface seismic and VSP data. Several deep neural networks based on three different architectures are developed and tested. The matrix propagator technique is used to generate the synthetic data for network training. The pre-trained deep neural networks can instantly predict velocity models from new data in a single step. The synthetic datasets used in training can be extended by adding random noise to the existing data, thus making the method closer to real-world conditions.
Paper
Full text
Seismic Inversion with Deep Neural Networks: a Feasibility Analysis
Semantic Scholar · Engineering · 2019
Abstract
Summary We investigate deep learning approaches to inversion of a 1D model of the subsurface using synthetic surface seismic and VSP data. Several deep neural networks based on three different architectures are developed and tested. The matrix propagator technique is used to generate the synthetic data for network training. The pre-trained deep neural networks can instantly predict velocity models from new data in a single step. The synthetic datasets used in training can be extended by adding random noise to the existing data, thus making the method closer to real-world conditions.