Near-Earth objects: a neural network calculates an asteroid’s orbit based on a single night’s observation

Trained on 568,127 arcs from 39,031 real-world objects, a physics-informed model proposes a plausible orbit in 85 per cent of cases, compared with around 45 per cent for the Gauss and Laplace methods.

When an asteroid close to Earth has just been discovered, there are often only a few positions recorded during a single night. With such a short arc, conventional methods of orbit determination become unstable or fail. Researchers propose to let a neural network directly estimate the distance and radial velocity of the object from three dated angular measurements, which completes its orbital state.

Two models compared

They compare a purely data-driven, multilayer perceptron and a variant with a physics-informed loss function, both trained on the 568,127 very short arcs of 39,031 NEOs in the NEODyS-2 catalog. Out of 56,773 test arcs, the physical model places 85.1% of its predictions in the dynamically permissible region, compared to 76.5% for the purely statistical version. The methods of Gauss and Laplace give a solution on only 45% and 47.3% of these arcs, and converge in about nine out of ten cases to a degenerate root.

The authors qualify the fact that physical stress is not better everywhere. It produces more extreme errors, but clearly outweighs fast, nearby objects, which are the most important for operational monitoring.