Beyond Correlations: Deep Learning for Seismic Interferometry.

Area of Science:

• Geophysics
• Seismology
• Machine Learning

Background:

• Passive seismic interferometry (SI) uses ambient noise to study Earth's structure.
• Traditional SI relies on cross-correlation, which has limitations due to source distribution and recording duration.
• Ergodicity assumptions for seismic sources are often unmet in real-world scenarios.

Purpose of the Study:

• To develop a novel method for passive seismic interferometry using deep neural networks (NNs).
• To overcome temporal and spatial limitations of correlation-based SI.
• To improve the accuracy of Green's function retrieval from passive seismic recordings.

Main Methods:

• Training deep neural networks (modified ResNet) within a supervised learning framework.
• Using correlograms (virtual shot gathers) as input and empirical Green's functions (active shot gathers) as desired output.
• Incorporating source directionality information (beamforming) to handle inhomogeneous source distributions.

Main Results:

• Deep NNs successfully retrieve Green's functions from short (5-min) passive seismic recordings with acceptable accuracy.
• The proposed NN approach overcomes limitations of traditional cross-correlation methods in SI.
• Synthetic examples demonstrate the retrieval of P-wave reflections at exploration scales.

Conclusions:

• Deep neural networks offer a powerful solution for advancing passive seismic interferometry.
• This method enhances the ability to extract valuable seismic information from ambient noise.
• The NN-based approach shows promise for improved crustal structure investigation and monitoring applications.