Machine Learning Approaches to Seismic Velocity Model and Seismogram Prediction in Earth’s Shallow Crust

Aug 4, 2026

Upload by Violet

Seismic imaging and subsurface property prediction are advanced through Fourier Neural Operators (FNOs), which can simulate seismic wavefields significantly faster than physics-based solvers after training. The work addresses the challenge of transferring simulation-trained performance to field-acquired seismic data, requiring representative small-scale velocity heterogeneities and topography. Using offshore Atlantic sedimentary basin surveys as a testbed, the study builds simulation-to-real workflows, training FNO models to predict velocity models and forward-modeled seismograms, with results so far indicating partial recovery of key arrivals and ongoing effort to reduce overfitting.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Overview

Aims and Objectives

Motivation and Dataset

Regional survey dataset (2013–2014)

Results In Progress: FNO Training

Title Machine Learning Approaches to Seismic Velocity Model and Seismogram Prediction in Earth’s Shallow Crust Creators Totten, Eoghan J. and Bean, Christopher J. and O'Brien, Gareth S. Date 2024 Citation Totten, Eoghan J. ORCID: [https://orcid.org/0000-0002-9958-1524](https://orcid.org/0000-0002-9958-1524) , Bean, Christopher J. ORCID: [https://orcid.org/0000-0003-3285-2446](https://orcid.org/0000-0003-3285-2446) and O'Brien, Gareth S. (2024) Machine Learning Approaches to Seismic Velocity Model and Seismogram Prediction in Earth’s Shallow Crust. In: Irish Geological Research Meeting, 1st-3rd March 2024, University of Galway. (Accepted Version) URL [https://dair.dias.ie/id/eprint/1416/](https://dair.dias.ie/id/eprint/1416/) Machine Learning Approaches to Seismic Velocity Model and Seismogram Prediction in Earth’s Shallow Crust E. Totten 1 , C. J. Bean 1 , G. S. O’Brien 2 1 Geophysics Section, Dublin Institute for Advanced Studies, 2 Microsoft Ireland, Dublin, Ireland 1. Overview Recent advances in machine learning present new ways for geoscientists to predict geological subsurface properties. Fourier Neural Operators (FNOs) [3] are increasingly being used as an alternative to conventional seismic imaging approaches [7] . FNOs have been shown to predict accurate simulations of seismic waves several hundred times faster than physics-based solvers post-training [2] . In synthetic volcanic settings, FNOs have been applied successfully to both the forward and inverse problem, capturing fine-scale velocity structure in heterogeneous models and seismograms [5] . However, transferring the successful performance of simulation-trained FNOs to field-gathered seismic data is yet to be attained. To achieve this, training models must contain representative small-scale velocity heterogeneities and topography to produce highly scattered codas in synthetic seismograms. This research presents work in progress on simulation-to-real FNO applications using field-gathered seismic data from offshore sedimentary basin settings as a testbed environment. Historical seismic survey datasets from Atlantic sedimentary basins are often accompanied by additional site-specific geological constraints. This makes the creation of synthetic velocity models and seismograms with field-derived properties possible, centering the collation of data for real-world machine learning applications in the numerical domain. The longterm research goal is to bring insights gained from training FNOson a better understood seismic environment to volcanic and other complex environments in future work. | | | | --- | --- | |

2. Aims and Objectives | | | Aim
● Train a Fourier Neural Operator (FNO) deep neural network to predict a seismic velocity model from a representative population of synthetic model: seismic record sets.
● Investigate how to populate seismic velocity models and data statistically. | Objective
● Focus FNO training on an extensively studied seismic setting with high quality data.
● Redeploy the FNO model in more challenging seismic environments (e.g. volcanoes), taking forward insights from more readily imaged settings. | 55oN 50oN | |

3. Motivation and Dataset

15oW

10oW

5oW | | | | --- | --- | --- | --- | |

Figure 1 .
Figure 1. Map of the 2013-2014 Regional Seismic Survey off the west coast of Ireland, eastern Atlantic Ocean. Each blue line represents a ship trajectory along which seismic profiling was performed. Figure credit: Department of the Environment, Climate and Communications.
This research requires a pilot dataset with the following requirements:
• High-quality and abundant seismic data
• Data recorded in a well-understood crustal study area with a priori geomorphological constraints.
Relative to volcanoes, sedimentary basins are more easily imaged as geological environments. The Porcupine Basin is an extensively studied structure with p

FAQs

What is the main machine learning approach discussed for seismic prediction?

The document focuses on Fourier Neural Operators (FNOs) to predict seismic wave-related outputs, including predicting seismograms from velocity models.

Why is simulation-to-real transfer challenging in this work?

Performance from simulation-trained FNOs does not automatically carry to field data; training must include representative small-scale velocity heterogeneities and topography to match real-world complexity.

What dataset environment is used as the testbed for the research?

Offshore sedimentary basin seismic survey data from Atlantic settings is used, including the 2013–2014 ENI-Regional Seismic Survey off the west coast of Ireland.

Aug 4, 2026

Upload by Violet