Neural representation-based waveform inversion via coupled foundational and generative networks
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Full waveform inversion (FWI) has served as a leading approach for seismic velocity model building, yet its highly nonlinear nature often hinders convergence and renders inaccurate solutions. Motivated by recent advances in deep learning,we aim to directly invert velocity from seismic data using a unified framework that handles both the seismic and velocity domains. Concretely, a pretrained foundational Transformer encoder extracts latent representations of the observed seismic waveforms, given by learned features. Meanwhile, an autoregressive Transformer decoder, dedicated to velocity synthesis,conditioned on these learned features, is used to to generate velocity models corresponding to the shot gathers. Through rigorous testing on homogeneous layered media and more intricate geological settings, the method demonstrates promising accuracy, highlighting its potential to capture complex subsurface features. The results suggest that our Transformer-based approach opens a new avenue for data-driven velocity inversion. We will show an application to realistic field scenarios in the presentation of this work.
Fields of Study & Subjects
Seismic Imaging and Inversion Techniques
Seismic Waves and Analysis
Model Reduction and Neural Networks
Publication details
- Journal / source
- Second International Meeting for Applied Geoscience & Energy
- Volume
- 1
- Pages
- 1164–1168
- Publication date
- 10 February 2026
- Language
- EN