← Back to Sessions

Akinet: A Physics Informed Neural Network for Building a Short-period Global Dispersion Model

Description: 

Dispersion measurements from ambient noise correlations are essential for structural imaging at various scales, from reservoirs to continents. However, extracting these measurements requires computationally intensive signal processing of large waveform datasets. Existing neural network methods, such as Disper-Net and DC-Net, focus on time-domain ambient noise correlations but have not addressed frequency-domain cross-spectra, particularly in cases of multi-component seismograms obtained from spatially and temporally discontinuous seismic arrays. We introduce AkiNet, a physics-informed neural network (PINN) designed to estimate phase velocity through waveform fitting using linear combinations of Bessel functions. AkiNet currently employs unsupervised learning to extract phase dispersion from low-quality ambient noise cross-spectra. By incorporating physics, observational data, and prior Earth model constraints, AkiNet enhances global coverage of short-period dispersion measurements while ensuring computational speed, efficiency, and accuracy. In inference mode, AkiNet will act as a generative model, enabling rapid termination of waveform fitting or dispersion modeling, particularly for multi-component cross-spectra. Preliminary results are promising, and ongoing work focuses on benchmarking AkiNet using the Africa-wide ADAMA dataset, optimizing model architecture, different learning methods, and refining regularization strategies. We anticipate that AkiNet will contribute to the development of global dispersion models, improving structural understanding of the Earth's crust and upper mantle lithosphere.

Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Siyu

Student Presenter: No

Invited Presentation: 

Poster Number: 140

Authors