A cyclic learning approach for improving pre-stack seismic processing.

Area of Science:

• Geophysics
• Machine Learning
• Data Science

Background:

• Seismic processing in oil and gas relies on expert interactions for data quality.
• Existing machine learning methods often overlook these expert-guided optimization processes.

Purpose of the Study:

• To present geocycles, a novel cyclic learning approach for pre-stack seismic processing.
• To integrate expert interaction principles into machine learning workflows for improved seismic data quality.

Main Methods:

• Developed a cyclic learning framework (geocycles) that refactors seismic processing into training, testing, and evaluation sub-tasks.
• Implemented greedy sequential sample selection for optimizing large seismic datasets.
• Applied the approach to noise attenuation and velocity analysis tasks.

Main Results:

• Geocycles demonstrated significant improvements in seismic data processing outcomes.
• Achieved up to 128% improvement in overall quality for two distinct seismic processing tasks compared to a single-cycle approach.
• Validated the effectiveness of cyclic structure and quality metrics in optimizing large datasets.

Conclusions:

• The geocycles approach effectively mimics expert iterative processes in seismic data optimization.
• Cyclic learning structures and robust quality metrics are crucial for enhancing seismic processing efficiency and data quality.