MasconCube: Fast and Accurate Gravity Modeling with an Explicit Representation

1. Problem Statement

The gravitational field modeling of irregularly shaped small bodies (asteroids and comets) faces significant challenges for deep space exploration missions (e.g., Hera, Psyche, Tianwen-2). Existing methodologies suffer from critical limitations:

• Spherical Harmonics: These diverge within the Brillouin sphere (the smallest sphere enclosing the body), making them unsuitable for spacecraft operating near the surface of irregular bodies.
• Polyhedral Models: While stable near the surface, they assume a homogeneous density distribution, failing to capture internal heterogeneity (porosity, voids, composition changes).
• Machine Learning Approaches (GeodesyNets, PINN-GM):
  • GeodesyNets: Use implicit neural representations. They are computationally expensive to train and lack direct physical interpretability.
  • PINN-GM: Incorporate physical constraints but remain computationally heavy and often fail to reconstruct accurate internal density profiles (converging to unrealistic central concentrations).
• General Issues: High computational costs, long training times, and the "ill-posed" nature of gravity inversion (non-unique solutions) make current methods inefficient for mission-critical applications requiring real-time adaptation or high-fidelity internal structure analysis.

2. Methodology: MasconCubes

The authors propose MasconCubes, a novel self-supervised learning approach that formulates gravity inversion as a direct optimization problem over a regular 3D grid of point masses (mascons).

Core Architecture

• Explicit Representation: Unlike implicit neural networks, MasconCubes explicitly model mass distributions on a fixed 3D Cartesian grid.
• Grid Initialization: A grid of $N^3$ mascons is defined around the target body. Mascons falling outside the known surface mesh (shape model) are removed, constraining the solution space to the physical volume.
• Optimization Target: The trainable parameters are the dimensionless masses ($m_j$) of the remaining mascons. The total mass is normalized to unity ($\sum m_j = 1$).
• Forward Model: Gravitational acceleration $\hat{g}_i$ at an observation point $r_i$ is calculated via Newtonian summation:
  $$\hat{g}_i = -G \sum_{j=1}^{\hat{N}} \frac{m_j(r_i - r_j)}{\|r_i - r_j\|^3}$$
• Loss Function: A normalized $L_1$ loss is used to minimize the discrepancy between predicted and observed accelerations. To handle global scale differences, an optimal scalar factor $c$ is introduced:
  $$L = \frac{1}{M} \sum_{i=1}^{M} \|g_i - c\hat{g}_i\|_1$$
• Training Strategy:
  • Self-Supervised: No manual labels are needed; the model learns from simulated or observed gravity data.
  • Shape Constraint: The method requires a priori knowledge of the asteroid's shape (surface mesh) to define the valid volume for mascons.
  • Efficiency: Training involves simple vector arithmetic and norm calculations, avoiding the iterative numerical integration required by implicit neural networks.

3. Key Contributions

1. Novel Formulation: Introduces a self-supervised learning framework that treats gravity inversion as a direct optimization of a 3D mascon grid, bridging the gap between physical mascon models and modern machine learning.
2. Computational Efficiency: Demonstrates training speeds approximately 40 times faster than GeodesyNets (reducing training time from ~290 minutes to ~6.6 minutes on an RTX 2080 Ti) while maintaining high accuracy.
3. Physical Interpretability: Provides explicit mass distributions, allowing mission planners to visualize internal density variations, unlike "black box" implicit neural representations.
4. Robust Evaluation: Establishes a rigorous benchmark using diverse asteroid models (Bennu, Eros, Itokawa, and synthetic planetesimals) with heterogeneous internal structures (layered, hollow, smooth gradients).

4. Results

The authors evaluated MasconCubes against GeodesyNets, PINN-GM III, and a "cubified" ground truth.

• Stokes Coefficients (Harmonic Analysis):
  • MasconCubes achieved Mean Absolute Errors (MAE) in normalized Stokes coefficients 1–2 orders of magnitude lower than GeodesyNets.
  • It accurately captured high-degree harmonic features essential for surface proximity operations.
• Acceleration Prediction:
  • MasconCubes showed superior accuracy in predicting gravitational acceleration vectors at various altitudes compared to GeodesyNets and uniform models.
  • Relative Euclidean norm errors were significantly lower across all tested altitudes.
• Internal Structure Reconstruction:
  • Heterogeneity: MasconCubes successfully reconstructed complex internal structures (e.g., layered densities in Bennu, smooth gradients in Itokawa).
  • PINN-GM Failure: PINN-GM III failed to reconstruct internal density, converging to a singular mass concentration at the center of mass, rendering it unsuitable for geodesy studies.
  • Hollow Bodies: Both MasconCubes and GeodesyNets struggled to reconstruct central voids (due to the Shell Theorem), but MasconCubes performed comparably to GeodesyNets in detecting asymmetries.
• Trajectory Simulation:
  • Spacecraft trajectory simulations using MasconCube-derived gravity fields resulted in significantly lower position errors compared to GeodesyNets and uniform models, validating its utility for navigation.

5. Significance and Limitations

Significance:

• Mission Critical: The combination of high accuracy, computational speed, and physical interpretability makes MasconCubes ideal for autonomous spacecraft operations, proximity navigation, and resource utilization planning.
• Paradigm Shift: It moves away from computationally expensive implicit neural networks toward efficient, explicit optimization, proving that machine learning can be highly effective for geodesy without sacrificing physical insight.

Limitations & Future Work:

• Shape Dependency: The current method requires a priori knowledge of the asteroid's shape. Training without shape constraints (shape-free) led to convergence on spherical distributions due to the ill-posed nature of gravity inversion from distant observations. Future work aims to develop regularization techniques to enable shape-free training.
• Internal Voids: Like all gravity-based inversion methods, reconstructing symmetric internal voids remains fundamentally difficult due to the non-uniqueness of the solution (Newton's Shell Theorem).

Conclusion:
MasconCubes represents a state-of-the-art advancement in asteroid geodesy, offering a fast, accurate, and physically interpretable solution for modeling the gravitational fields of irregular small bodies, directly addressing the needs of next-generation deep space exploration missions.