Scaling uncertainties on asteroid characteristics to prepare datasets for machine learning

Abstract

Physical and dynamical characterisations of asteroids are used in different fields, such as Solar System formation modelling, Planetary Defence and Resources Prospecting. The vast majority of asteroids are not known in detail - have at best their orbit well defined - and the knowledge on the composition or internal structure is derived by models of reflectivity curves, with limited certainty. Machine learning methods have begun to be used on asteroid datasets, but the major uncertainties about their characteristics are slowing down the applicability. This paper reviews some stakes and challenges of asteroid exploration, and why the introduction of common characterisation factors would be beneficial for the asteroid science community, especially with the application of machine learning methodologies. A preliminary scale to quantify the characterisation of asteroids is proposed, and finally discusses its interests and limitations for machine learning applications. The investigated characteristics of asteroids are: size/shape, orbital dynamics, mass/density, spin, internal structure and composition. This paper reviews the current methods used to determine these parameters, and provides a preliminary scale based on the certainty associated with the different measurements. Characterisation factors are useful to build datasets that will be used in machine learning algorithms applied to asteroid science. The ratio of currently known asteroids in each defined bin of characterisation factor is estimated. Moreover, a total characterisation factor that yield a preliminary quantification of our knowledge about a specific asteroid (i.e. the sum of all the certainties about its different characteristics) is defined. Finally, characterisation factors for specific applications can be introduced using an adapted weighting system. This preliminary work provides a baseline for scaling uncertainties of asteroids properties. The next step is to create viable datasets using application specific characterisation factors that would allow the use of advanced machine learning algorithms already available.