Search the EPA Research Database

Project title

Rapid Retrieval of Seismic Velocity Models Using Fourier Neural Operator-based Deep Learning

Primary Funding Agency

Irish Research Council

Co-Funding Organisation(s)

Lead Organisation

Dublin Institute for Advanced Studies

Lead Applicant

Not listed

Project Abstract

With the exponential growth of seismic data available for study in the past decade, there is a need to develop seismic images and estimate velocity models in quicker and more computationally efficient ways than current inversion techniques. Moreover, the climate emergency in 2022 highlights the need to reduce the environmental impact of seismic imaging. Machine Learning has accelerated the ability of seismologists to relate seismic waveforms to the geological properties of the subsurface. With sufficient training, neural networks can infer seismic velocity maps, at a fraction of the computational cost of conventional inversion-based methods. Current state-of-the-art algorithms replace convolutional neural networks (CNN) with Fourier Neural Operators (FNOs) due to their computational efficiency. In addition to replacing expensive convolution operations with simple multiplication, FNOs take low frequency seismic data as input, and still predict successful outputs at the required resolution, provided that waveforms and velocity model populations used during training have statistical distributions representative of the �ground truth�. One outstanding research question is: how close do synthetic datasets need to represent real-world ground truth, to maximise insights from neural networks? This project investigates general approaches to populate training data for FNOs, by creating families of synthetic seismic waveforms in a hybridised statistical-deterministic approach. The objective is to improve neural network performance, so fewer real-time observations are needed to recover seismic models. This reduces the number of active seismic sources required during surveys, shifting collation of massive data to the numerical domain. This research focuses first on sedimentary basin settings, before tackling complex volcanic geological settings. Neural network-driven recovery of subsurface models has the potential to transform real-time environmental hazard advisories for landslides, volcanic eruptions and more. It creates potential for less invasive, less carbon-intensive, more ethical seismic image construction and geo-engineering, a pillar of the global push for net zero emissions.

Grant Approved

�105,604.00

Research Hub

Climate Change

Research Theme

3. Climate Solutions, Transition Management and Opportunities

Start Date

01/09/2023

Initial Projected Completion Date

31/08/2023