2D magnetotelluric forward modeling based on multitask deep learning.

Area of Science:

• Geophysics
• Computational Science
• Artificial Intelligence

Background:

• Accurate 2D magnetotelluric (MT) forward modeling is crucial for geophysical inversion quality.
• Traditional numerical methods are computationally intensive, limiting their efficiency on personal computers.

Purpose of the Study:

• To develop a novel and efficient 2D MT forward modeling method using a deep learning approach.
• To accelerate MT forward calculations while maintaining high accuracy.

Main Methods:

• A Transformer U-Net (T-Unet) multitask network was employed for end-to-end training.
• The network learns the mapping between geoelectric models and apparent resistivity/phase data.
• A trained neural network model directly predicts MT forward modeling results.

Main Results:

• The T-Unet model significantly reduces computation time compared to traditional simulations.
• High computational accuracy is maintained after the model establishment.
• The deep learning method demonstrates superior efficiency on personal computers.

Conclusions:

• Deep learning neural networks hold significant potential for accelerating MT forward calculations.
• This research offers a new avenue for integrating artificial intelligence in geophysical exploration.
• The T-Unet method provides an efficient and accurate alternative for 2D MT forward modeling.