Asteroid shape inversion with light curves using deep learning

TL;DR

This work addresses the slow and local-minima-prone nature of traditional asteroid shape inversion from light curves by proposing a deep-learning pipeline that directly maps photometric observations to a convex-hull 3D point cloud. The method employs a two-stage architecture with a geometry-aware transformer to predict coarse and refined shapes, trained on large simulated datasets, achieving millisecond-level performance. It further introduces a concavity-detection module that leverages the difference between convex-hull and non-convex light curves to identify concave regions, with IoU scores reaching up to 0.89 for large features. Validations on simulated data and real Lowell Observatory observations demonstrate robustness and potential for onboard autonomous decision-making, offering a practical route toward rapid, scalable asteroid shape estimation and non-convex region identification.

Abstract

Asteroid shape inversion using photometric data has been a key area of study in planetary science and astronomical research.However, the current methods for asteroid shape inversion require extensive iterative calculations, making the process time-consuming and prone to becoming stuck in local optima. We directly established a mapping between photometric data and shape distribution through deep neural networks. In addition, we used 3D point clouds to represent asteroid shapes and utilized the deviation between the light curves of non-convex asteroids and their convex hulls to predict the concave areas of non-convex asteroids. We compared the results of different shape models using the Chamfer distance between traditional methods and ours and found that our method performs better, especially when handling special shapes. For the detection of concave areas on the convex hull, the intersection over union (IoU) of our predictions reached 0.89. We further validated this method using observational data from the Lowell Observatory to predict the convex shapes of the asteroids 3337 Milo and 1289 Kuta, and conducted light curve fitting experiments. The experimental results demonstrated the robustness and adaptability of the method

Figures (20)

• Figure 1: Illustration of our proposed inversion method. The top four panels show the schematic diagrams of different visible and illuminated facets of the asteroid under varying Sun-asteroid-Earth positional relationships in the asteroid's body coordinate system. The blue point cloud represents the asteroid point cloud obtained from the inversion process.
• Figure 2: Network architecture. First, FPS is used to downsample the input light curve sampling points in order to determine the center points. Then, the nearest neighbor set corresponding to each center point is obtained with the KNN algorithm, and local features are extracted using a simplified 1D-DGCNN. Afterwards, positional embeddings are added to enhance spatial awareness. The enhanced features are encoded by a geometry-aware transformer module followed by global feature aggregation through max pooling. Finally, the coarse point cloud is refined into the output using a transformer decoder that integrates three up-sampling layers and skip connections.
• Figure 3: Method of predicting non-convex regions.
• Figure 4: Simulated light curve for convex.
• Figure 5: Part of non-convex 3D model datasets.