Magnetotelluric forward modeling on fine grid via deep learning with physical information constraints

Area of Science:

• Geophysics
• Computational Science
• Artificial Intelligence

Background:

• Traditional magnetotelluric (MT) forward modeling faces efficiency bottlenecks with fine grids due to computational complexity.
• Deep learning (DL) approaches for MT modeling often compromise physical accuracy and struggle to utilize fine grid precision.

Purpose of the Study:

• To develop a high-precision, high-efficiency MT forward modeling solution by synergizing DL with traditional methods.
• To address the limitations of existing MT forward modeling techniques in handling complex subsurface properties and computational demands.

Main Methods:

• Generated realistic synthetic resistivity models using cubic spline interpolation to capture subsurface volume effects and gradual variations.
• Developed and trained a U-shaped deep learning model (Swin-UNet with Swin Transformer backbone) under strict physical constraints for multi-task MT forward response grid refinement.

Main Results:

• The integrated framework effectively utilizes fine grids, achieving high precision in MT forward modeling.
• Demonstrated a significant reduction in forward modeling time compared to traditional methods.
• Validated the model's ability to retain physical information and fit real MT data accurately.

Conclusions:

• The study presents an efficient and high-precision solution for MT forward modeling, overcoming traditional computational bottlenecks.
• This work pioneers the application of AI in fine grid forward modeling, offering a new paradigm for geophysical modeling and interpretation.
• Provides a valuable reference for interdisciplinary geophysical modeling and data interpretation.