Joint 3D Gravity and Magnetic Inversion via Rectified Flow and Ginzburg-Landau Guidance

This paper introduces a novel framework for 3D gravity and magnetic joint inversion that reframes the problem as a rectified flow on the Noddyverse dataset, utilizing a Ginzburg-Landau regularizer and guidance methodology to capture solution distributions and improve ore identification while also releasing a trained VAE for 3D densities.

The Gist

Imagine you are a detective trying to solve a mystery, but you can't see the crime scene. You only have two clues left behind on the surface of the ground: a map of gravity (how heavy the ground feels) and a map of magnetism (how magnetic the ground is).

Your goal? To figure out exactly what the "treasure" (ore deposits) looks like deep underground.

The problem is that this is a magic trick gone wrong. Many different shapes of treasure could create the exact same gravity and magnetism maps on the surface. This is called an "ill-posed problem." It's like trying to guess the shape of a hidden object just by looking at its shadow; a round ball and a flat disk could cast the same shadow from a specific angle.

Here is how the authors of this paper solved it, using a mix of modern AI and old-school physics.

1. The Old Way vs. The New Way

• The Old Way (Classical Math): Traditional methods try to find one perfect answer. They guess a shape, check if it fits the clues, and tweak it until it works. But because there are so many possible answers, they often get stuck on just one "average" guess that might look like a blurry blob rather than a sharp rock.
• The New Way (AI & Probability): Instead of guessing one answer, this paper teaches an AI to imagine thousands of possible answers. It doesn't just say, "The ore is here." It says, "There is a 90% chance the ore is a sharp, jagged rock here, and a 10% chance it's a smooth sphere there." This gives geologists a much better picture of the uncertainty.

2. The "Noddyverse" Training Gym

To teach the AI, you need practice data. You can't just use real rocks because you don't know what's actually underground (that's the mystery!).

• The Analogy: Imagine a video game simulator called Noddyverse. It's a physics engine that builds millions of fake underground worlds with known treasures.
• The researchers trained their AI on this simulator. The AI learned to look at the "shadows" (gravity/magnetism) and instantly recognize the shapes of the hidden rocks. It's like a student who has studied millions of practice exams before taking the real test.

3. The Secret Sauce: "Rectified Flow"

The AI uses a technique called Rectified Flow.

• The Analogy: Think of a messy room full of scattered toys (noise). You want to clean it up to reveal a perfect toy castle (the ore).
• Most AI methods try to clean the room by taking tiny, random steps, which can be slow and wobbly.
• Rectified Flow is like having a magical vacuum cleaner that knows the exact straight line to pull the toys from the mess directly into the castle. It's faster and more efficient, allowing the AI to generate high-quality 3D models quickly.

4. The "Ginzburg-Landau" Guide (The Physics Teacher)

This is the most clever part. Even with a great AI, sometimes the generated rocks look too smooth or "blobby." Real ore deposits usually have sharp, distinct boundaries between the valuable rock and the useless dirt.

• The Problem: The AI might try to make a rock that is half-ore and half-dirt, which doesn't exist in nature.
• The Solution: The authors added a "Physics Teacher" called the Ginzburg-Landau (GL) regularizer.
• The Analogy: Imagine you are molding clay. The AI wants to make a smooth, round ball. The GL teacher is a strict sculptor who says, "No! Rocks have sharp edges. You must separate the clay from the dirt clearly."
• The GL theory acts like a force that pushes the AI to create sharp, clean lines between different materials, mimicking how nature actually works.

5. How They Used It (The "Plug-and-Play" Trick)

The researchers didn't just train a new AI from scratch. They built a plug-and-play module.

• The Analogy: Imagine you have a generic robot that can draw anything. You want it to draw only realistic rocks. Instead of retraining the whole robot, you just clip a "Rock Guide" onto its arm.
• This "Guide" (the GL method) nudges the robot's drawing in real-time. If the robot starts drawing a blurry blob, the Guide pushes it to sharpen the edges. If it draws a shape that doesn't match the gravity clues, the Guide pushes it back.

The Result

When they tested this new system:

1. It was faster than old methods.
2. It was more accurate, producing 3D models that looked like real geological structures (sharp boundaries, distinct layers).
3. It handled uncertainty better, giving geologists a range of likely scenarios rather than a single, potentially wrong guess.

In a Nutshell

The authors built a super-smart AI detective trained on a physics video game. They gave it a straight-line shortcut (Rectified Flow) to solve the mystery faster and added a strict physics teacher (Ginzburg-Landau) to ensure the solution looked like a real rock, not a blurry cloud. This helps miners find valuable resources more accurately without having to dig up the whole planet.

Technical Summary

1. Problem Statement

Geophysical Inversion Challenge:
The core problem is the joint 3D inversion of gravitational and magnetic surface data to reconstruct subsurface distributions of density ($\rho$) and magnetic susceptibility ($\chi$).

• Ill-posedness: The problem is fundamentally ill-posed and non-unique; many distinct subsurface models can produce identical surface measurements.
• Limitations of Current Methods:
  • Deterministic Algorithms: Traditional methods (e.g., Tikhonov regularization) converge to a single "best-fit" solution, failing to capture the distribution of possible geological scenarios or quantify uncertainty.
  • Standard Machine Learning: Most ML approaches predict a single point estimate without modeling the full posterior distribution. Furthermore, many are trained on simplistic data that lacks real-world geological complexity.
  • Probabilistic Methods: While Monte Carlo methods offer probabilistic sampling, they are computationally prohibitive for high-dimensional 3D models.
• Specific Geological Constraint: Ore deposits typically exhibit sharp boundaries between the ore body and the host rock (phase separation), a feature difficult to capture with smooth regularization but essential for accurate identification.

2. Methodology

The authors propose a framework that reframes the inversion problem as posterior sampling using a Rectified Flow generative model, enhanced by Ginzburg-Landau (GL) physics.

A. Dataset and Pre-processing

• Noddyverse: The model is trained on the Noddyverse dataset, the largest physics-based synthetic dataset for geophysical inversion. It contains 3D blocks ($200 \times 200 \times 200$ voxels) of density and susceptibility, paired with forward-simulated 2D surface gravity and magnetic fields.
• Pipeline: A custom pipeline handles extraction, Zarr-based compression for efficient I/O, and log-transform normalization to handle the heavy-tailed distribution of magnetic susceptibility.

B. Model Architecture

1. 3D Variational Autoencoder (VAE):
   • To handle the high dimensionality ($200^3$ voxels), the authors train a 3D VAE to compress the joint density-susceptibility volumes into a lower-dimensional latent space ($24^3$).
   • The VAE uses residual convolutions and bottleneck attention mechanisms.
2. Rectified Flow Model:
   • Instead of standard diffusion, the authors use Rectified Flow, which learns a straight-line trajectory between noise and data distributions, enabling faster convergence and sampling.
   • The model operates in the VAE's latent space, predicting the flow velocity field.
   • Two backbones were tested: a 3D UNet and BiFlowNet (a hybrid UNet/Transformer architecture), with BiFlowNet showing superior efficiency.

C. Physics-Aware Regularization (Ginzburg-Landau)

To enforce the sharp boundaries typical of ore deposits, the authors introduce the Ginzburg-Landau (GL) functional, a continuous relaxation of the discrete Ising model.

• GL Energy: Defined as $\mathcal{E}_{GL}[\phi] = \int (\frac{\kappa}{2}|\nabla \phi|^2 + W(\phi)) dx$, where $W(\phi)$ is a double-well potential favoring states $\phi = \pm 1$ (representing host rock vs. ore).
• Integration Strategies:
  1. GL-Regularized Training: The GL energy is computed on latent grids and used to reweight the training loss, encouraging the model to learn phase-separation dynamics.
  2. GL Guidance (Inference): A novel "plug-and-play" guidance step is applied during the reverse flow process. The score predicted by the flow model is augmented by the gradient of the GL energy:
     $$s_{combined} = s_{\theta} + \lambda_{GL}(t) \cdot (-\nabla \mathcal{E}_{GL})$$
     This steers the sampling trajectory toward physically plausible, sharp-boundary solutions without retraining the model.

D. Inversion Algorithm (FlowDPS)

The inversion uses Diffusion Posterior Sampling (DPS) adapted for Rectified Flow in the latent space:

1. Forward Operator: The VAE decoder and the physical forward operator (gravity/magnetic kernels) are made differentiable.
2. Data Consistency: At each step of the reverse flow, the latent estimate is refined to minimize the mismatch between the predicted surface data and the actual observations.
3. Guidance: The GL guidance term is applied to the refined latent to enforce geological realism (sharp interfaces).

3. Key Contributions

1. First Generative Flow for Joint Inversion: This is the first work to develop a generative model specifically for joint magnetic and gravitational inversion, enabling stochastic reconstruction of 3D property models.
2. GL Physics for Ore Identification: Introduction of a Ginzburg-Landau regularizer that encodes phase-separation structure, effectively modeling the sharp boundaries between ore and host rock in a continuous, differentiable manner.
3. Plug-and-Play GL Guidance: A methodology to integrate physics-based priors (GL) into existing unconditional denoisers (Rectified Flow) as a modular guidance step, avoiding the need for complex retraining.
4. Community Assets: Release of a trained 3D VAE for 3D densities/susceptibilities and a pre-processing pipeline for the Noddyverse dataset, facilitating future research in the field.

4. Results

The authors evaluated three variants: (i) Baseline Flow, (ii) Flow with GL Regularization (training-time), and (iii) Flow with GL Guidance (inference-time).

• Qualitative: GL Guidance produced 3D volumes with geologically plausible structures and sharp boundaries that were visually consistent with observed fields.
• Quantitative:
  • Error Reduction: GL Guidance consistently achieved lower Root Mean Square Error (RMSE) compared to the baseline and the GL-regularized training variant for both magnetic and gravity fields.
  • Distribution: Empirical Cumulative Distribution Functions (ECDF) showed that GL Guidance shifted the error distribution significantly toward positive improvements ($\Delta RMSE > 0$).
  • Win Rates: GL Guidance achieved the highest "best-rate" and lowest "worst-rate" across 128 generated samples.
• Insight: The authors note that GL Guidance outperformed GL Regularization, hypothesizing that applying regularization directly in the latent space (during training) may not perfectly capture the complex physical dynamics observed in the decoded susceptibility space.

5. Significance

• Uncertainty Quantification: By moving from deterministic point estimates to posterior sampling, this method provides a distribution of possible subsurface models, crucial for risk assessment in mineral exploration.
• Scalability: The combination of VAEs and Rectified Flow makes high-dimensional 3D inversion computationally feasible, overcoming the bottlenecks of traditional Monte Carlo methods.
• Physics-Informed AI: The work demonstrates how deep generative models can be effectively constrained by physical laws (via GL theory) to produce geologically realistic results, bridging the gap between data-driven AI and geophysical physics.
• Open Science: The release of code, models, and datasets lowers the barrier to entry for researchers in geophysical inversion and generative AI.