What type of study is this?

This is a Quantitative Study study.

September 17, 2025

Prediction of Near-Surface Velocity from Seismic Traveltimes Data Using Machine Learning

Key Points

The novel framework predicts near-surface velocity with high accuracy, improving seismic data correction processes.
Using a deep neural network, the method provides consistent results comparing predicted thicknesses and velocities to real-world data.
The machine learning-based inversion streamlines the process, eliminating significant computational burdens associated with traditional techniques.
Training on synthetic datasets allows the approach to manage nonlinearity and dependence on initial models effectively.

Abstract

Abstract Accurate determination of near-surface velocity is essential for calculating static corrections for seismic data, as neglecting shallow velocity can lead to significant misplacement of deeper seismic reflectors. However, traditional methods for inverting refracted traveltimes data are often computationally intensive and time-consuming. To overcome these limitations, we present a novel machine learning-based inversion framework that predicts the thickness and velocity of near-surface layers from refracted traveltimes data. Our approach utilizes a data-driven methodology to approximate the relationship between seismic traveltimes and corresponding velocity and thickness, reducing ambiguities in seismic inversion results, particularly without sufficient prior information. We trained a deep neural network (DNN) model using synthetic traveltimes datasets and evaluated its performance on synthetic and real-world data. The results demonstrate that the proposed approach can accurately estimate layer thicknesses and velocities with high consistency between predicted and actual values. Furthermore, applying the machine learning approach to a field dataset acquired over a dune in Saudi Arabia, comprising 37 receivers at 2 m intervals, yielded results consistent with borehole data. The proposed machine learning framework offers several advantages over traditional inversion techniques, including automatically identifying valuable inversion information without prior knowledge, avoiding non-uniqueness challenges, and efficient training on smaller datasets without extensive computational resources. Additionally, the approach can manage nonlinearity associated with seismic inversion and reduce dependence on initial models, providing automatic solutions in milliseconds. Although this approach depends on the quality and amount of training data, it marks a notable step forward in seismic inversion and offers substantial potential to enhance the accuracy and efficiency of static seismic corrections.

Perguntar à IA

Bookmark