About half the stars in the Milky Way live in pairs. In the dense cores of globular clusters, such binaries frequently bump into a third star. These rendezvous last mere days or weeks, but the consequences are enormous: they heat up the cluster, preventing it from collapsing, and create unusual objects — from blue stragglers to tight pairs of black holes and neutron stars, whose mergers are detected as gravitational waves. The trouble is that the gravitational three-body problem is chaotic: a tiny change in initial conditions leads to a completely different outcome — for instance, the interloper might kick out one partner and take its place, or all three may fly apart forever.
Astrophysicists compiled a dataset of nearly a million encounters using the REBOUND code, varying masses, orbits, velocities, and impact angles over wide ranges. Each simulation was classified into one of four outcomes: ionization (the pair breaks apart), flyby (the single star exits without altering the binary), exchange (the intruder replaces one companion), and hierarchical triple system (the third star stays on a wide orbit). Then they balanced the classes and trained an Machine Learning model, XGBoost, based on decision trees. Its input features were not raw parameters but physically meaningful dimensionless quantities: binary hardness (the ratio of binding energy to the kinetic energy of the incoming star), focusing factor, the Safronov number, and others. This approach allowed the model to capture nonlinear patterns.
On a held-out set of 198 000 simulations, prediction accuracy reached 88.32%, and top‑2 accuracy (the correct class is among the two most likely) hit 98.77%. The dominant predictor turned out to be hardness: a tightly bound binary tends to survive the encounter or exchange components; a loosely bound one falls apart. The model is well calibrated: its probability estimates closely match the actual frequencies of outcomes. Errors are not random: they concentrate at the boundary between exchange and hierarchical triples. In this narrow parameter region, energy and momentum are redistributed so delicately that the stars temporarily capture one another in a chaotic resonance, and the result hinges on the tiniest details. Basin entropy analysis confirmed the fractal structure of this boundary: exactly here, not elsewhere, the model wavers most often and delivers low confidence.
This result turns the expensive three-body calculation (minutes per event) into something nearly instantaneous (milliseconds). The model can be plugged into star cluster simulations—like MOCCA or CMC—where direct integration of every encounter would eat up centuries of computer time. Now, dynamical effects including resonances and triple formation can be reproduced faithfully, not approximately.
The next step is to include effects that were neglected so far: tidal forces, finite stellar sizes, and relativistic corrections. This is crucial for predicting mergers of compact objects. Moreover, the model can be extended to predict not just the outcome class but also the final orbits, velocities, and lifetimes.
The method will impact population studies of gravitational-wave sources and our understanding of cluster evolution as a whole.
Incorporate calibrated probabilities into hybrid schemes: doubtful cases are computed exactly, confident ones via the surrogate.
The work directly ties to the classical three-body problem, whose chaotic nature was recognized by Henri Poincaré in the late 19th century. It demonstrates that physical chaos harbors hidden structures accessible to machine learning—a step toward predictability in fundamentally unpredictable systems.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The work trains an XGBoost machine learning model on nearly a million three-body simulations to instantly predict the outcome of a binary star encountering a single star. The input features are physically meaningful dimensionless quantities, with the main predictor being the hardness of the binary system. The model achieves 88% accuracy and is well calibrated, with errors concentrated in a narrow chaotic region at the boundary between exchange and hierarchical triple systems.
Where its strength is
The shift from expensive calculations to near-instant predictions while preserving accuracy, including resonances and the formation of triple systems, opens the possibility of embedding the model into star cluster simulations and accurately reproducing dynamical effects.
Where it could go next
Worth trying
Nearby works in our archive
The closest work in the archive sits at 0.66 by meaning — the usual level for our corpus.
In some configurations, three stars form a temporary hierarchical tandem that drifts through the cluster for millions of years before breaking apart.
In Liu Cixin’s famous novel 'The Three-Body Problem,' the Trisolaran civilization suffers from the unpredictable orbits of three suns. Astronomers face a similar problem at the scale of star clusters, and now some of that unpredictability has been tamed by an algorithm.