ARTFEED — Contemporary Art Intelligence

Unsupervised Depth-Aware Neural Networks for 3D Gravity Inversion

other · 2026-08-11

A recent arXiv preprint (2608.08959) introduces an unsupervised depth-aware implicit neural representation aimed at 3D gravity inversion. Gravimetry captures subsurface density variations linked to geological formations, geothermal systems, and intrusive bodies. The challenge of deriving a three-dimensional density model from gravity data is significant due to issues like non-uniqueness, sparse data coverage, and gravity field attenuation with depth. Traditional inversion techniques depend on explicit regularization and parameter adjustments, while supervised deep learning methods often lack sufficient gravity-density pairs. The new approach utilizes multiple coordinate-based neural networks for overlapping depth slabs, optimized using the sensitivity matrix derived from observed gravity data. It incorporates slab-specific Fourier features, physics-informed depth gains, and scheduled regularization to establish structural priors, all in an unsupervised manner, eliminating the need for labeled data.

Key facts

  • Paper arXiv:2608.08959v1 proposes unsupervised depth-aware implicit neural representation for 3D gravity inversion.
  • Gravimetry images subsurface density contrasts associated with geological structures, geothermal systems, and intrusive bodies.
  • 3D density model recovery from gravity observations is ill-posed due to non-uniqueness, limited data coverage, and depth attenuation.
  • Classical inversion methods rely on explicit regularization and parameter tuning.
  • Supervised deep-learning approaches require representative gravity–density pairs that are rarely available.
  • The density volume is represented by multiple coordinate-based neural networks assigned to overlapping depth slabs.
  • Optimization is done directly from observed gravity measurements through the sensitivity matrix.
  • Slab-specific Fourier features, physics-based depth gains, and scheduled regularization provide structural priors.

Entities

Institutions

  • arXiv

Sources