Accelerated Full Waveform Inversion by Deep Compressed Learning.

Area of Science:

• Geophysics
• Computational Science
• Machine Learning

Background:

• Full Waveform Inversion (FWI) requires massive datasets, posing computational challenges for complex subsurface modeling.
• Industrial-scale seismic data can reach teraflop storage levels, making extensive FWI analysis prohibitive.

Purpose of the Study:

• To develop a method for reducing the dimensionality of FWI inputs to mitigate computational costs.
• To accelerate large-scale 3D FWI by optimizing data selection.

Main Methods:

• A deep neural network with a binarized sensing layer learns optimal seismic acquisition layouts through compressed learning.
• Representation learning using an autoencoder generates latent representations of shot gathers.
• K-means clustering is applied to latent representations for hierarchical selection of relevant shot gathers.

Main Results:

• The proposed method consistently outperforms random data sampling in FWI.
• Effective data reduction was achieved, utilizing only 10% of the data for 2D FWI with superior results.
• The approach demonstrates potential for accelerating large-scale 3D inversion.

Conclusions:

• The developed deep learning technique offers a viable solution for computational cost reduction in FWI.
• This method paves the way for more accessible and efficient large-scale subsurface imaging and analysis.