EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

807. Seismic optimizing inversion with multi-head attention mechanism and transformer

Invited abstract in session WC-42: Nonlinear optimization algorithms and applications, stream Variational Analysis and Continuous Optimization.

Wednesday, 12:30-14:00
Room: 98 (building: 306)

Authors (first author is the speaker)

1. Yanfei Wang
Institute of Geology and Geophysics, Chinese Academy of Sciences

Abstract

The technology of reflection coefficient inversion aims to broaden the spectrum and increase the dominant frequency based on the effective collection of data. Reflection coefficient inversion can be achieved through conventional physics-based optimization modeling. In recent years, machine learning has seen tremendous development in the field of inverse problems. Deep learning has made breakthrough progress in fields such as image processing and natural language processing (NLP). Following the convolutional neural network (CNN) and recurrent neural network (RNN), the transformer network architecture, comprised only of self-attention mechanisms and feedforward neural networks, has also become a research hotspot in recent years. Our latest work has realized blind seismic reflection coefficient inversion using a transformer network. By setting up appropriate encoders and decoders and incorporating physical mechanism constraints, the transformer has demonstrated powerful seismic inversion capabilities. Experimental data indicates that the new method based on the transformer network architecture does not require any a priori assumptions about the seismic records’ reflection coefficient series, seismic wavelets, or Q value models; it possesses strong noise resistance and resolution; after the model training is completed, it has relatively good generalization performance, and the model prediction process is more efficient compared to traditional methods.

Keywords

Status: accepted

Back to the list of papers