Rayleigh-wave dispersion data selection and model fine-tuning based on uncertainty estimation.

Area of Science:

• Geophysics
• Seismology
• Machine Learning

Background:

• Rayleigh-wave inversion is crucial for subsurface shear-wave velocity (SWV) determination, vital for seismic risk assessment, resource exploration, and geotechnical engineering.
• Current deep learning (DL) approaches for Rayleigh-wave inversion face challenges including limited generalization, heavy reliance on training data, and slow convergence.

Purpose of the Study:

• To develop an improved deep learning strategy for Rayleigh-wave inversion that addresses limitations of existing methods.
• To enhance the accuracy, robustness, and adaptability of DL models for SWV structure determination.

Main Methods:

• A representative data selection strategy identifying high-uncertainty samples using predictions from multiple parallel pretrained models.
• An automatic differentiation-driven inversion method to generate high-confidence pseudo-labels for selected data.
• Fine-tuning the original model with the generated pseudo-labels, a process independent of borehole information.

Main Results:

• Significant improvements in prediction accuracy and model robustness in the target area.
• Validation through both synthetic and field experiments, confirming enhanced performance.
• Demonstrated adaptability in complex geological environments with minimal additional cost.

Conclusions:

• The proposed data selection and model optimization strategy effectively enhances DL-based Rayleigh-wave inversion.
• The method offers a robust and adaptable solution for SWV structure analysis, particularly in challenging geological settings.
• This approach provides a cost-effective way to improve seismic data interpretation without requiring borehole data.