Seismic Velocity Model Building with Deep Convolutional Neural Networks SEP-190 (2023)
Downloads
Table of contents
- Chapter 1: Introduction
- Chapter 2: Wave-equation and deep learning theory
- Chapter 3: Earth model building with supervised learning
- Chapter 4: Tiber WATS Velocity Model Building
- Appendix A: Elastic wave equation as a linear system
- Appendix B: Acoustic wave equation as a linear system
- Bibliography
Abstract
This thesis focuses on recovering wave velocity estimates from seismic data with limited recording offsets and band-limited frequency content using Deep Learning (DL), specifically using deep convolutional neural networks (CNN). A combined strategy combines a supervised learning framework to recover the low-wavenumber components of velocity models with the traditional Full Waveform Inversion (FWI) to recover high-wavenumbers. We demonstrate the methodology on synthetic industry benchmark problems and a field data application with an open-source Gulf of Mexico seismic dataset.
Reproducibility and source codes
This thesis has been tested for reproducibility. The source codes and training data are made available at these GitHub repositories:
https://premonition.stanford.edu/sfarris/vmb-net.
https://premonition.stanford.edu/sfarris/vmb-net-data.