Learning-Based Seismic Velocity Inversion with Synthetic and Field Data.

Area of Science:

• Geophysics
• Machine Learning
• Subsurface Modeling

Background:

• Accurate acoustic subsurface velocity models are crucial for industrial exploration.
• Traditional inversion methods face challenges in complex geological regions.
• Deep learning (DL) offers a potential alternative but requires robust validation with field data.

Purpose of the Study:

• To analyze deep learning's capability for velocity model recovery using field and synthetic seismic data.
• To evaluate the impact of training data selection and augmentation on DL model performance.
• To assess DL's effectiveness in challenging geological settings like the Gulf of Mexico.

Main Methods:

• Utilized labeled field-recorded and synthetically generated seismograms for training DL models.
• Evaluated model performance using quantitative metrics (MSE, SSIM, R2).
• Assessed the geological plausibility and impact on geophysical migration images.

Main Results:

• Models trained on field data outperformed those trained on synthetic data across all metrics.
• Field-data trained models produced more geologically plausible results and sharper migration images.
• Synthetic data models showed potential but required advanced techniques to bridge the domain gap.

Conclusions:

• Deep learning, particularly with field-recorded seismograms, can significantly advance velocity model-building workflows.
• Synthetic data can be a viable alternative when field data is limited, provided domain expertise is applied.
• Earth scientists' expertise is vital for curating effective synthetic datasets.