MasconCube: Fast and Accurate Gravity Modeling with an Explicit Representation

Pietro Fanti 1, Dario Izzo 1 \cor · 2025

Read MasconCube: Fast and Accurate Gravity Modeling with an Explicit Representation by Pietro Fanti 1, Dario Izzo 1 \cor (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-08607.txt.

Introduction

Mapping gravity around asteroids and comets is the goal of Pietro Fanti and 1 co-author in a 2025 paper introducing MasconCube. Spacecraft fly close to these irregular bodies, where older spherical models fail. Polyhedral models assume even density, and neural methods can be slow to train.

MasconCube places point masses, called mascons, on a regular three-dimensional grid and learns their values directly. Known shape data limits the possible answers. Tests on Bennu, Eros, Itokawa, and synthetic bodies beat rival methods on several measures. Training runs about forty times faster than GeodesyNets, while the mass layout stays easy to read.

Catalog record

CREATORPietro Fanti 1, Dario Izzo 1 \cor
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-08607

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MasconCube: Fast and Accurate Gravity Modeling with an Explicit Representation Pietro Fanti 1, Dario Izzo 1 \cor Abstract The geodesy of irregularly shaped small bodies presents fundamental challenges for gravitational field modeling, particularly as deep space exploration missions increasingly target asteroids and comets. Traditional approaches suffer from critical limitations: spherical harmonics diverge within the Brillouin sphere where spacecraft typically operate, polyhedral models assume unrealistic homogeneous density distributions, and existing machine learning methods like GeodesyNets and Physics-Informed Neural Networks (PINN-GM) require extensive computational resources and training time. This work introduces 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). Unlike implicit neural representations, MasconCubes explicitly model mass distributions while leveraging known asteroid shape information to constrain the solution space. Comprehensive evaluation on diverse asteroid models including Bennu, Eros, Itokawa, and synthetic planetesimals demonstrates that MasconCubes achieve superior performance across multiple metrics. Most notably, MasconCubes demonstrate computational efficiency advantages with training times approximately 40 times faster than GeodesyNets while maintaining physical interpretability through explicit mass distributions. These results establish MasconCubes as a promising approach for mission-critical gravitational modeling applications requiring high accuracy, computational efficiency, and physical insight into internal mass distributions of irregular celestial bodies. \cor [email protected] 1. Advanced Concepts Team, European Space Research and Technology Centre (ESTEC), Noordwijk, The Netherlands. Keywords Gravity Models \cdot Mascon Models \cdot Asteroid Geodesy \cdot Machine Learning 1 Introduction The geodesy of irregularly shaped small bodies, particularly asteroids and comets, presents fundamental challenges that have driven decades of research in gravitational field modeling. As deep space exploration missions increasingly target these celestial objects, from previous missions like NEAR-Shoemaker [ 1 , 2 ] and Hayabusa [ 3 , 4 , 5 ] to current endeavors such as Hera [ 6 ] , Psyche [ 7 ] , and Tianwen-2 [ 8 ] , accurate characterization of their internal mass distributions has become essential for mission planning, proximity operations, and scientific understanding. The problem of inferring internal density distributions from external gravity measurements is inherently ill-posed, presenting significant computational and mathematical challenges [ 9 ] . Unlike large planetary bodies with relatively regular shapes, asteroids and comets exhibit highly irregular geometries that complicate traditional gravitational modeling approaches. The sparse and noisy nature of gravity measurements, combined with the non-unique solutions inherent to gravity inversion, makes the reconstruction of internal mass distributions particularly challenging. Furthermore, the computational complexity [ 10 ] required to achieve high-fidelity models while maintaining numerical stability near these irregular surfaces compounds these difficulties. Spherical harmonics have long served as the foundation for gravitational field representation [ 11 , 12 , 13 ] , offering analytical solutions to Laplace’s equation and providing a mathematically elegant framework for potential field description. However, their application to irregular small bodies faces critical spectral limitations [ 14 , 15 , 16 ] . Polyhedral gravity models [ 17 , 18 , 19 ] have emerged as a popular alternative, providing analytical solutions that remain stable within the Brillouin sphere and down to the body’s surface. However, they carry a fundamental limitation: they assume a homogeneous density distribution throughout the body. This assumption may not accurately reflect the internal composition of real asteroids, which can exhibit significant density variations due to porosity, composition changes, or internal voids. Recent advances in machine learning have introduced GeodesyNets [ 20 , 21 ] , neural networks that learn three-dimensional density distributions as continuous functions. These networks represent the density field as a neural implicit function, allowing for both shape reconstruction and internal structure inference without requiring prior shape knowledge. GeodesyNets demonstrate remarkable capability in capturing complex gravitational fields with relative errors below 1\% even near asteroid surfaces. However, the training process can be computationally intensive, particularly for complex heterogeneous models, and requires several data points. Physics-Informed Neural Networks for Gravity Modeling (PINN-GM) [ 22 , 23 ] , with their latest iteration PINN-GM III [ 24 ] , represent another machine learning approach that incorporates physical constraints directly into neural network training to learn gravitational potential. These models demonstrate advantages in compactness, noise robustness, and sample efficiency, but remain computationally expensive to train. Although PINN-GM shows promise for gravity modeling, their capability for geodesy reconstruction has been speculated, but never empirically tested. Current approaches exhibit several critical limitations that motivate the development of alternative methodologies. Training efficiency remains a persistent challenge across machine learning approaches, with GeodesyNets and PINN-GM models often requiring extensive computational resources and training time to achieve convergence. The spherical harmonics fidelity problem, where traditional analytical methods fail to capture fine-scale gravitational features of irregular bodies, continues to limit mission-critical applications requiring high accuracy near surfaces. Reproducibility concerns arise from the complex hyperparameter spaces and initialization sensitivities of neural network approaches, making it difficult to consistently achieve optimal performance across different types of asteroid and mission scenarios. Additionally, existing methods often struggle with the trade-off between computational efficiency and physical accuracy. While mascon models offer direct physical interpretation through point mass distributions, they traditionally require large numbers of masses to achieve fine-grained accuracy, leading to computational bottlenecks. The lack of a unified approach that combines the physical interpretability of mascon methods with the representational power of modern machine learning techniques represents a significant gap in current capabilities. The MasconCube approach introduced in this work addresses these limitations by formulating gravity inversion as a direct optimization problem over a regular 3D grid of point masses, leveraging self-supervised learning to reconstruct internal mass distributions efficiently while maintaining physical interpretability and computational tractability. 2 Methodology Figure 1 : Simplified scheme of the proposed method. A 3D mesh is employed to remove points from a randomly initialized MasconCube. The resulting mascon model predicts an acceleration \mathbf{\hat{g}} at an input position \mathbf{r} , which is then compared with the reference value \mathbf{g} . The computed loss is backpropagated to update the masses M . This section presents the comprehensive methodology employed to develop, train, and evaluate MasconCubes for asteroid gravity modeling. We begin by describing the diverse suite of simulated asteroid models constructed for benchmarking. We then detail the training approach used to optimize the mass distribution within the 3D mascon grid, including the observation model, loss function, and optimization procedures. Finally, we outline the multi-faceted evaluation framework that assesses model performance. 2.1 Asteroid Models To rigorously evaluate the performance of MasconCubes and benchmark it against GeodesyNets and PINN-GM III, we constructed a suite of simulated asteroid models derived mostly from well-characterized real-world bodies. Our selection criteria emphasized diversity in shape complexity, size, and scientific relevance, ensuring that our experiments capture a representative range of geodetic challenges encountered in planetary science. We selected a fictitious planetesimal body previously used in the literature [ 20 , 24 ] , along with three asteroids for which high-resolution shape models are publicly available: 101955 Bennu, 433 Eros, and 25143 Itokawa. For consistency with prior studies [ 20 ] , we employed simplified, low-resolution versions of their surface meshes. To operate in a dimensionless framework, each mesh was scaled to fit within a sphere of radius 0.8. The resulting volumes were then discretized into tetrahedra, with a mass concentration (mascon) assigned to each one,

Read the complete text