Inversion of Magnetic Data using Learned Dictionaries and Scale Space

Technical Summary: Inversion of Magnetic Data using Learned Dictionaries and Scale Space

Problem Statement
Magnetic data inversion is a fundamental tool in geophysics for inferring subsurface magnetic susceptibility distributions from surface measurements. However, the problem is inherently ill-posed, characterized by non-unique solutions, depth ambiguity, and high sensitivity to noise. Traditional inversion methods rely on pre-determined regularization schemes (e.g., $\ell_1$-$\ell_2$ norms, total variation) to stabilize solutions. While effective, these fixed regularization strategies often lack the adaptability required to capture complex, diverse geological structures. Furthermore, classical approaches typically iterate to full convergence, which can lead to overfitting in noisy environments.

Methodology
The authors propose a data-driven framework that integrates learned dictionaries with a scale-space approach to address the limitations of fixed regularization. The core methodology involves the following components:

1. Synthesis-Based Regularization: Instead of analyzing the model directly, the method assumes the subsurface model $m$ can be synthesized from a sparse latent vector $z$ via a dictionary operator $\Psi$ ($m = \Psi z$). The inversion seeks to find a sparse $z$ that minimizes the data misfit and the $\ell_1$ norm of the coefficients.
2. Learned Dictionaries: Rather than using fixed bases (e.g., wavelets or cosine transforms), the dictionary $\Psi$ is learned from a training set of plausible geological models. The dictionary is formulated as a set of 3D convolutional operations (a 4D tensor).
3. Scale-Space Framework (Unrolling): The optimization process is "unrolled" over a fixed number of iterations ($N$). This mimics an iterative soft-thresholding algorithm where the solution is progressively refined.
   • Shared Dictionary: A single dictionary $\Psi$ is used across all iterations (similar to a standard deep learning layer).
   • Unrolled (Dynamic) Dictionary: A distinct dictionary $\Psi^{(j)}$ is assigned to each iteration $j$. This allows the regularization to evolve dynamically, introducing structural details progressively without being rooted in a single global variational functional.
4. Training Procedure: The system is trained using synthetic data generated from a geostatistical distribution of geological models (specifically simulating porphyry deposits). The loss function combines the model reconstruction error ($\| \hat{m} - m \|^2$) and the forward data misfit ($\| A\hat{m} - d \|^2$). Stochastic gradient descent with automatic differentiation is used to update the dictionary weights.

Key Contributions

• Integration of Scale-Space and Dictionary Learning: The paper extends classical variational inversion by incorporating multi-scale representations through a scale-space framework, enabling the progressive introduction of structural detail.
• Dynamic Regularization: The authors introduce an "unrolled" dictionary learning technique where iteration-dependent dictionaries provide greater flexibility compared to fixed or shared dictionaries. This approach treats the regularization as a dynamical process rather than a static constraint.
• Deep Learning Analogy: The work establishes a direct connection between the proposed iterative inversion method and deep convolutional neural networks, where the dictionary acts as convolutional layers and soft-thresholding acts as the non-linearity.

Results
The method was evaluated on a synthetic 3D dataset ($64 \times 64 \times 32$ grid) simulating mineral deposits with 1% relative noise added to the forward data. Four methods were compared:

1. Variational approach (gradient-based).
2. Sparse recovery with a fixed cosine dictionary.
3. Sparse recovery with a shared learned dictionary.
4. Sparse recovery with an unrolled (dynamic) learned dictionary.

Quantitative Findings:

• The unrolled dictionary achieved the best model recovery, with a Normalized Mean Squared Error (nMSE) of 0.289, representing a significant improvement over the variational baseline (0.692) and the fixed cosine dictionary (1.000).
• The shared learned dictionary (0.625) showed improvement over the variational method but was outperformed by the unrolled approach.
• Data recovery was also superior for the learned methods, with the unrolled dictionary achieving an nMSE of 0.021 compared to 0.000 for the variational method (which perfectly fits noise in the training set context) and 0.983 for the fixed dictionary.

Qualitative Findings:
Visual comparisons of reconstructed 3D models demonstrated that the unrolled dictionary approach captured features deeper in the subsurface with higher fidelity to the true model compared to other methods. The scale-space approach effectively mitigated overfitting while recovering complex structural details.

Significance and Claims
The paper claims that the proposed data-driven approach significantly improves reconstruction accuracy and robustness compared to conventional variational and fixed-dictionary methods. The authors emphasize that the combination of learned dictionaries with scale-space dynamics offers a promising alternative for standard dictionary learning in geophysical inversion.

The work highlights the potential of this approach for applications in geophysical exploration, environmental assessment, and mineral prospecting. However, the authors maintain a modest stance regarding generalization, noting that the framework requires a training dataset derived from geological modeling. They acknowledge that further research is needed to understand the limitations and generalization properties of the proposed regularization when applied to models that are out-of-distribution. The code for the study is made publicly available to facilitate reproducibility.