Recent advances in modeling density distributions, so-called neural density fields, can accurately
describe the density distribution of celestial bodies without, e.g., requiring a shape model - properties of great advantage when designing trajectories in close proximity to these bodies. Previous
work introduced this approach, but several open questions remained. This work investigates neural density fields in the context of robustness to external factors like noise or constraints during
training, like the maximal available gravity signal strength due to a certain distance. Thus, this
work studies the relative errors of 2824 trained neural networks for the two bodies 433 Eros and
67P/Churyumov-Gerasimenko with varying training conditions. Models trained with mascon or
the polyhedral ground truth without any noise perform similarly in the experiments with relative
errors below 5% until 100 m distance for training in the near-range. The impact of solar radiation
pressure on a typical probe affects training neglectably, with the relative error being of the same
magnitude as without noise. Limiting the training point accuracy to 1 mGal severely hurts the
obtainable precision when only sampling above several kilometers of height. However, inducing
a relative error of 10% stills leads to acceptable results, with errors below 5% in the sampling
range used for training and more than 50 times away. Finally, pretraining is shown as practical in
order to speed up network training. Hence, geodesyNet is a helpful tool extending the toolbox of
assembling a navigation model for a small body.
All code and results are publicly available at https://github.com/gomezzz/geodesyNets
«
Recent advances in modeling density distributions, so-called neural density fields, can accurately
describe the density distribution of celestial bodies without, e.g., requiring a shape model - properties of great advantage when designing trajectories in close proximity to these bodies. Previous
work introduced this approach, but several open questions remained. This work investigates neural density fields in the context of robustness to external factors like noise or constraints during
train...
»