Deep neural networks for analysis of Mercury’s planetary exosphere

Abstract

The emergence of Artificial Intelligence’s Deep Neural Networks (DNNs) as a method for analysis of spatial and temporal data gives us a new avenue for research of the processes inherent to planetary exospheres and their interaction with the planetary environment. Hereby, it will be presented a particular study on how in-situ observations of the elemental composition of Mercury’s exosphere may serve as an indication of the surface regolith mineral composition below, through predictions given by pre-trained DNNs. In this case the main driver, considered to generate the exosphere, is the micrometeoroid impact vaporization (MIV), which is seen as the dominant process in the night-side hemisphere of the planet. The training datasets will be constructed from randomly generated virtual surfaces, which will aim to generalize the training of the AI models to various types and compositions of the regolith. Different Neural Network models, which include fully-connected networks and Convolutional Neural Networks, will be compared both in giving supervised classification through multivariate regression predictions, and in reconstructing the regolith mineralogy maps. Furthermore, ways to expand the Deep Neural Networks to pattern recognition and knowledge discovery will be explored beyond the surface-exosphere interaction, as well as the possibilities for transfer learning and online learning with the acquisition of real planetary data. The development of such data analysis algorithms will be shown especially in view of the upcoming arrival of the ESA/JAXA’s BepiColombo mission to Mercury, the innermost planet, in 2025, with its variety of instruments able to capture the dynamics of the tenuous Hermean environment. Ultimately, would be targeted the implementation of such AI algorithms within the Ground Segment pipeline software architecture of the experiment BepiColombo/MPO/SERENA, composed by four ion and neutral particle detectors.