Convolutional Neural-Network-Based Reverse-Time Migration with Multiple Reflections.

Area of Science:

• Geophysics
• Computational Seismology
• Machine Learning in Earth Sciences

Background:

• Reverse-time migration (RTM) provides high-resolution seismic images but is limited by velocity model accuracy, aperture illumination, and computational efficiency.
• Least-squares RTM (LSRTM) improves reflectivity but remains sensitive to velocity model accuracy.
• RTM with multiple reflections (RTMM) enhances illumination but introduces crosstalk artifacts.

Purpose of the Study:

• To develop a novel method for enhancing the quality of seismic images generated by RTMM.
• To address the limitations of RTM and RTMM, specifically concerning resolution, accuracy, and computational cost.
• To leverage convolutional neural networks (CNNs) for improved seismic data processing.

Main Methods:

• A residual U-Net based convolutional neural network (CNN) was designed to act as a filter, learning the inverse Hessian.
• The CNN was trained to map reflectivity from RTMM to true reflectivity, utilizing an identity mapping for residual learning.
• The trained network, termed RTMM-CNN, was applied to enhance RTMM-generated seismic images.

Main Results:

• RTMM-CNN successfully recovered major subsurface structures and thin layers with enhanced resolution and accuracy compared to standard RTM-CNN.
• The method demonstrated significant generalizability across diverse geological models, including complex thin layers, salt bodies, folds, and faults.
• RTMM-CNN exhibited lower computational cost than LSRTM, indicating improved computational efficiency.

Conclusions:

• The proposed RTMM-CNN method effectively enhances seismic image quality by mitigating artifacts and improving resolution.
• This approach offers a computationally efficient and generalizable solution for complex subsurface imaging challenges.
• RTMM-CNN represents a significant advancement in applying machine learning to seismic migration techniques.