EclipseNETs: Learning Irregular Small Celestial Body Silhouettes
Giacomo Acciarini, Dario Izzo, Francesco Biscani · 2025
Read EclipseNETs: Learning Irregular Small Celestial Body Silhouettes by Giacomo Acciarini, Dario Izzo, Francesco Biscani (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-2504-04455.txt.
Introduction
Giacomo Acciarini, Dario Izzo, and Francesco Biscani teach, in a 2025 paper, neural networks to hold the silhouette of an asteroid, so eclipse forecasts around it become nearly instant. A spacecraft near a small body needs to know when it will slip into shadow, for power and heat planning. Tracing rays against a detailed shape model is exact but slow, and small bodies are all lumps and facets.
Their networks store the outline implicitly and answer eclipse questions orders of magnitude faster than ray tracing, at comparable accuracy on four famous bodies: Bennu, Itokawa, the comet 67P, and Eros. A second framework learns the shape from sparse flight path data alone, with no prior model, refining the forecast as new passes arrive. The authors aim the method at missions that must navigate rubble piles they have never seen up close.
Catalog record
| CREATOR | Giacomo Acciarini, Dario Izzo, Francesco Biscani |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-04455
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
EclipseNETs: Learning Irregular Small Celestial Body Silhouettes Giacomo Acciarini, Dario Izzo, Francesco Biscani Abstract Accurately predicting eclipse events around irregular small bodies is crucial for spacecraft navigation, orbit determination, and spacecraft systems management. This paper introduces a novel approach leveraging neural implicit representations to model eclipse conditions efficiently and reliably. We propose neural network architectures that capture the complex silhouettes of asteroids and comets with high precision. Tested on four well-characterized bodies - Bennu, Itokawa, 67P/Churyumov-Gerasimenko, and Eros - our method achieves accuracy comparable to traditional ray-tracing techniques while offering orders of magnitude faster performance. Additionally, we develop an indirect learning framework that trains these models directly from sparse trajectory data using Neural Ordinary Differential Equations, removing the requirement to have prior knowledge of an accurate shape model. This approach allows for the continuous refinement of eclipse predictions, progressively reducing errors and improving accuracy as new trajectory data is incorporated. K eywords Small Bodies \cdot Silhouette Reconstruction \cdot Asteroids \cdot Comets \cdot NeuralODE \cdot Neural Events \cdot Machine Learning \cdot Artificial Intelligence \cdot AI for Space \cdot Spaceflight Mechanics \cdot Orbital Dynamics 1 Introduction Small celestial bodies, including asteroids and comets, play a critical role in space exploration. They provide insights into the early Solar System, offer potential resources for future missions, and pose impact threats to Earth [ 1 , 2 ] . As remnants of Solar System formation, they preserve primordial materials that inform our understanding of early chemical and physical conditions. Additionally, they present unique opportunities for scientific exploration, such as sample-return missions and spacecraft technology testing [ 3 , 4 , 5 , 6 ] . As their exploration becomes more advanced, understanding the dynamics of spacecraft as they orbit or interact with these bodies is critical, particularly given their irregular shapes, weak gravity, and complex physical properties [ 7 ] . Precise eclipse condition estimation is vital in small-body exploration, impacting solar radiation pressure calculations, communication, and spacecraft thermal and power management [ 8 , 9 , 10 ] . For objects with high area-to-mass ratios orbiting weak-gravity bodies, solar radiation pressure is a dominant perturbation, making accurate eclipse computation essential for predicting spacecraft motion [ 11 , 12 , 13 ] . Eclipse computation involves determining whether an orbiting object’s position lies within a celestial body’s shadow [ 14 ] . Shadow modeling depends on the body’s shape and size; simple bodies allow for approximations using cylindrical cones, while irregular bodies require high-fidelity 3D models [ 15 ] . For example, the NASA OSIRIS-REx mission employed a detailed 3D model of asteroid 101955 Bennu to simulate the precise shadow region and the motion of ejected particles [ 16 , 13 ] . Traditional eclipse modeling methods, such as the Möller–Trumbore intersection algorithm [ 17 ] , use ray-tracing techniques to determine shadow regions. While highly accurate, these methods are computationally expensive and non-differentiable, limiting their utility in reliably estimating eclipse entrance and exit conditions [ 18 ] . Recent advances in neural image processing introduce differentiable and computationally efficient alternatives using implicit neural representations, particularly Neural Radiance Fields (NeRFs). These models capture complex 3D scenes with photometric consistency [ 19 ] . Periodic activation functions, as explored by Sitzmann et al. [ 20 ] , enhance the ability to model high-frequency details, making them well-suited for accurately describing shadows cast by irregular celestial bodies [ 18 , 21 ] . In this paper, we construct a database of irregular silhouettes for four small celestial bodies using their shape models. The dataset is split into training and validation sets to train neural networks, termed EclipseNETs, for reconstructing small-body silhouettes from different directions. We first compare periodic activation functions with ReLU-based architectures confirming their superior ability in capturing complex shapes. Then, we integrate these models into orbital propagators to assess their accuracy and computational efficiency relative to traditional methods. The results demonstrate that EclipseNETs can efficiently and accurately perform eclipse condition computation. Due to the differentiable nature of EclipseNETs, this approach further facilitates the use of Taylor propagators [ 18 ] , enhancing the reliability and accuracy of eclipse event detection during orbital propagation. In the second part of the paper, we introduce a training technique based on Neural Ordinary Differential Equations (NeuralODEs) [ 22 , 23 ] , which enables the direct learning of silhouettes from trajectory data. This approach eliminates the need for precise prior knowledge of the irregular shapes of small bodies, which are often not accurately known before in situ exploration. By incorporating EclipseNET, the resulting spacecraft dynamics can be formulated as a NeuralODE, allowing direct optimization of the network parameters from trajectory data. This enables learning and/or refinement of the eclipse model by minimizing trajectory prediction errors, presenting a novel strategy to improve irregular eclipse descriptions through NeuralODEs. The paper is structured as follows: Section 2 discusses methods and related work, focusing on eclipse function definitions, the irregular bodies studied, and the spacecraft dynamics. Section 3 details the training process of EclipseNETs, comparing Siren-based and ReLU-based architectures. We also discuss EclipseNET integration in a Taylor-based orbital propagator, benchmarking it against traditional ray-intersection methods. Section 4 introduces an algorithm for direct silhouette learning from trajectory data via NeuralODEs. Finally, Section 5 summarizes our findings. 2 Methods 2.1 Eclipse Function An eclipse function, F_{\mathcal{B}}(x,y,\hat{\mathbf{s}}) , is introduced to implicitly represent the eclipsed region created by an irregular body \mathcal{B} when illuminated by a point source at infinity in the direction \hat{\mathbf{s}} (in the body frame). This function is defined in a reference frame lying in a plane orthogonal to \hat{\mathbf{s}} . Mathematically, the eclipse function is given by: F_{\mathcal{B}}(x,y,\hat{\mathbf{s}})=\left\{\begin{array}[]{l}d(x,y,\partial\Omega_{\hat{\mathbf{s}}})\ \textrm{(outside the eclipse region)}\\ -d(x,y,\partial\Omega_{\hat{\mathbf{s}}})\ \textrm{(inside the eclipse region)}\\ \end{array}\right. where d(x,y,\partial\Omega_{\hat{\mathbf{s}}}) represents the signed distance function from the point ( x,y ) to the boundary \partial\Omega_{\hat{\mathbf{s}}} of the eclipse region. This function is closely related to the signed distance function, commonly used in computer graphics [ 24 ] , but here it describes two-dimensional shapes parametrized by the projection direction \hat{\mathbf{s}} . The gradient of the eclipse function in the direction perpendicular to the eclipse boundary satisfies: \nabla_{x,y}F_{\mathcal{B}}\cdot\hat{\mathbf{n}}=1\text{,} where \hat{\mathbf{n}} is the unit normal (either outward or inward) to \partial\Omega_{\hat{\mathbf{s}}} . Figure 1 provides a visual representation of the eclipse function and its relevant geometric properties. While Figure 2 B) illustrates examples of the eclipse function for a specific direction across different bodies. Figure 1: Schematic illustration of the shadow cast by an irregular body, and its eclipse function F_{\mathcal{B}} . 2.2 Irregular Bodies Our focus is on four different small bodies: 101955 Bennu, 67P / Churyumov - Gerasimenko, 433 Eros, and 25143 Itokawa. In all cases, high-fidelity polyhedral models reconstructed from various instruments on board spacecraft that visited these celestial bodies are utilized. For Eros and Itokawa, models produced by Robert Gaskell [ 25 , 26 ] are used; for Bennu, the model provided by the OSIRIS-REx team [ 27 ] is employed; and for Churyumov–Gerasimenko, the model available from the European Space Agency [ 28 ] is used. Figure 2: A) 3D models of Bennu, Churyumov-Gerasimenko, Eros, and Itokawa. B) Contour plot of the eclipse function for a fixed viewpoint. C) Sampled points of the eclipse function used to build the