Accuracy in orbital simulations depends on a single energy-angular-momentum balance, not error size.

Simulating objects like colliding neutron stars is difficult. Computers must track both fast orbits and slow energy loss over time without letting tiny math errors snowball into huge timing mistakes. To solve this, researchers built three kinds of neural flows called SINFONIA. A neural flow is an AI model designed to predict how a physical system changes over time while respecting core physical laws. Instead of taking tiny steps, the model learns a step forward in time and applies it repeatedly across an entire spiral toward collision. The researchers found that long-term accuracy depends on controlling error along one specific path dictated by energy and angular momentum balance. By protecting this balance, the models ran over 10^{2} to 10^{5} steps—even with steps lasting a full orbit or more—beating a standard solver with lower cost and smaller phase errors. The models could even discover missing physical forces.

Why it matters

Gravitational-wave scientists need to simulate star orbits over long times without errors piling up. This approach makes simulations both faster and more accurate, and it can even reveal hidden forces in the data.

Questions answered