01 / AI · Analysis
ML benchmark for chaotic three-body orbits publishes dataset
Stability prediction and long-horizon trajectory prediction differed in difficulty, the researchers wrote; the abstract gives no accuracy gaps between models

A machine learning benchmark for chaotic three-body orbits appeared on arXiv, a public preprint server, on 9 Nov 2026, with two main tasks and its code and dataset public. The paper's abstract names those tasks as stability prediction and trajectory prediction.
The abstract's main conclusion concerns the order of difficulty. The researchers wrote that, within the same chaotic system, judging stability and predicting an exact trajectory differed in difficulty.
The argument starts with the trajectories the researchers used as data.
Early dynamics and stability prediction
The researchers said they first simulated the trajectories of triple systems, groups of three bodies moving under mutual gravity, from initial conditions near the critical boundary between stable and unstable. Stability prediction is the task of classifying whether a triple system has diverged up to a set horizon.
The abstract reports that "early dynamics enable strong stability prediction". In plain terms, stability classification performed well when it used a system's early motion.
The paper said the comparison set includes the Mardling-Aarseth stability criterion from astrophysics. The machine learning models range from simple regression to state-of-the-art foundational physics models. The abstract does not give the accuracy gap between the criterion and each model.
The difficulty of long-horizon trajectories
Trajectory prediction takes input features and returns the exact trajectory, the abstract explained. Unlike stability prediction, which asks only whether a system diverged, it requires the trajectory itself to be right.
The researchers wrote that their experiments revealed a clear order of task difficulty in chaotic three-body dynamics. Stability classification using early dynamics performed strongly, while precise long-horizon trajectory prediction remained difficult. The abstract does not state what length of time counts as long-horizon.
Reproduction passes to the public code
According to the abstract, judging divergence and predicting exact trajectories differed in difficulty in this benchmark. The result covers astrophysical three-body systems, and the abstract contains no game engine experiment.
The researchers said they released the code and dataset in a GitHub repository named ThreeBodyMLBench. The release allows other models to be run on the same tasks and their results compared.