Summary
This technical report presents a hybrid machine learning (ML) workflow designed for dual-sensor towed-streamer data, addressing three computationally intensive stages in marine seismic preprocessing: deblending, denoising, and deghosting. The report outlines how each ML step minimizes manual parameter tuning, with denoising models trained on pseudo-synthetic examples derived from real noise and cleaned records. The deblending process employs an ML-generated seed to expedite sparse iterative inversion, while the deghosting model integrates hydrophone and geophone inputs using a modified DuckNet architecture. Validation results indicate that the dual-sensor approach significantly enhances model performance, achieving a validation loss of 0.017, compared to 0.025 for a hydrophone-only model. The workflow is applied to a 3D survey along the Equatorial Margin of Brazil, demonstrating a substantial reduction in processing time for individual sail lines, which can now be completed within hours. The findings suggest that this ML-accelerated approach effectively reduces the manual effort typically required in traditional seismic data processing.
