Asteroid Albedo and Size Inversion: A Deep Learning Method and Application Based on Multiparameters

Abstract

Of nearly 1.5 million asteroids discovered to date, only about 5% have spectroscopic observations, resulting in limited physical property coverage. Multicolor photometry provides taxonomic classifications for the majority of asteroids. Meanwhile, asteroids originating from the same collisional event and occupying nearby regions of orbital space typically share similar surface properties. Consequently, orbital parameters can provide statistical constraints on asteroid type and albedo, enabling more reliable estimates of their sizes. To achieve accurate inversion of asteroid size and albedo using readily available information under optical observation and to improve the completeness of asteroid physical property databases, this paper proposes a deep learning method that takes an 11-dimensional vector comprising absolute magnitude, orbital parameters, and type information as input, and embeds the Bowell formula to fuse physical priors with data-driven modeling. The proposed method eliminates the dependence on spectral data, thereby significantly expanding the training sample size, which in turn ensures robust model generalization. Test results on asteroids explored by spacecraft or radar observations show mean absolute percentage errors of 26.4% and 21.6% for albedo and effective diameter, respectively. By applying the model to all asteroids with known types, a large-scale asteroid physical properties catalog was constructed. The resulting catalog contains taxonomic type, geometric albedo, and effective diameter for 188,737 asteroids. Compared to existing catalogs based on machine learning predictions, the sample coverage has improved by an order of magnitude. The generated catalog is available on Zenodo (doi:10.5281/zenodo.20032293).