Implicit neural representation for potential field geophysics.

Area of Science:

• Geophysics
• Artificial Intelligence
• Data Science

Background:

• Potential field geophysics traditionally relies on gridding methods.
• Multilayer perceptron (MLP) neural networks with spatial coordinates offer new possibilities.
• Implicit neural representations can model complex data functions.

Purpose of the Study:

• To present a novel method for implicit neural representation of potential fields.
• To encode and evaluate geophysical survey data using coordinate MLP networks.
• To compare the neural network approach with traditional gridding techniques.

Main Methods:

• Utilizing coordinate multilayer perceptron (MLP) networks for implicit function learning.
• Encoding synthetic and real airborne geophysical survey data.
• Employing automatic differentiation for gradient calculation within the neural network framework.

Main Results:

• The proposed method generated a regular grid closely matching synthetic forward models (10.3 nT RMSE).
• Horizontal gradients calculated via the neural network were accurate compared to numerical methods.
• The neural network approach demonstrated rapid training using single survey extent data.

Conclusions:

• Coordinate MLP networks provide an effective method for potential field data representation.
• This approach offers advantages in accuracy and computational efficiency over traditional methods.
• Further research may improve vertical gradient calculations for specific datasets.