Seismic Waveform Inversion Capability on Resource-Constrained Edge Devices.

Area of Science:

• Geophysics
• Computational Seismology
• Artificial Intelligence

Background:

• Seismic full wave inversion (FWI) is crucial for subsurface imaging but is computationally intensive and requires significant human input.
• Deep learning methods offer acceleration but often face challenges with overfitting and stability.

Purpose of the Study:

• To develop an edge computing-based, data-driven seismic inversion technique using supervised deep convolutional neural networks.
• To enable accurate and efficient reconstruction of subsurface velocities on resource-constrained devices.

Main Methods:

• Trained deep convolutional neural networks (UNet and InversionNet) on raw seismic data and velocity models to learn the data-velocity mapping.
• Deployed and executed the trained models on an edge device (Raspberry Pi) for real-time inference.
• Evaluated model robustness against noise through noise-aware and no-noise training strategies.

Main Results:

• Achieved comparable inference times on Raspberry Pi (2-22 seconds) versus GPU (2-18 seconds) for UNet and InversionNet models.
• Demonstrated robustness to noise, enabling reliable velocity model estimation under various signal-to-noise ratios.
• Developed a user-friendly GUI for automated inversion on the edge device.

Conclusions:

• The proposed edge computing-based deep learning approach significantly accelerates seismic full wave inversion.
• Real-time, on-device seismic data processing is feasible on resource-constrained edge devices.
• The technique provides a robust and efficient solution for subsurface velocity imaging, even in the presence of noise.