Three-dimensional inversion of gravity data using implicit neural representations and scientific machine learning.

Area of Science:

• Geophysics
• Machine Learning
• Computational Science

Background:

• Gravity inversion is crucial for understanding subsurface density variations.
• Applications include mineral exploration, geothermal energy, carbon storage, and groundwater management.
• Current methods often rely on mesh-based discretizations, which can limit scalability and flexibility.

Purpose of the Study:

• To present a novel scientific machine learning approach for 3D gravity inversion.
• To represent subsurface density as a continuous field using implicit neural representations (INRs).
• To overcome limitations of traditional mesh-based methods.

Main Methods:

• Utilized a deep neural network trained via a physics-based forward-model loss.
• Employed spatial encoding to improve the capture of sharp density contrasts and short-wavelength features.
• Applied the method to synthetic datasets, including complex geological models and dipping block structures.

Main Results:

• The INR framework successfully reconstructed detailed subsurface structures and geologically plausible boundaries.
• Achieved accurate inversion without explicit regularization or depth weighting.
• Demonstrated reduced inversion parameters with increasing problem size, indicating scalability.

Conclusions:

• Implicit neural representations offer a scalable, flexible, and interpretable framework for large-scale geophysical inversion.
• This approach shows significant potential for advancing gravity inversion techniques.
• The framework is adaptable for other geophysical methods and joint/multiphysics inversions.