A machine learns to predict the stability of circumbinary planets

TL;DR

This paper introduces a deep neural network that predicts the stability of circumbinary planets more accurately than traditional methods, enabling faster and more reliable assessments in planetary system studies.

Contribution

The authors develop and validate a DNN model trained on N-body simulations to improve stability predictions for circumbinary planets, surpassing existing polynomial criteria.

Findings

DNN achieves at least 86% accuracy near stability boundaries.

The model outperforms traditional polynomial criteria in capturing resonant islands.

The DNN is fast enough for real-time Bayesian inference applications.

Abstract

Long-period circumbinary planets appear to be as common as those orbiting single stars and have been found to frequently have orbital radii just beyond the critical distance for dynamical stability. Assessing the stability is typically done either through N-body simulations or using the stability criterion first considered by Dvorak and later developed by Holman and Wiegert: a second-order polynomial calibrated to broadly match numerical simulations. However, the polynomial is unable to capture islands of instability introduced by mean motion resonances, causing the accuracy of the criterion to approach that of a random coin-toss when close to the boundary. We show how a deep neural network (DNN) trained on N-body simulations generated with REBOUND is able to significantly improve stability predictions for circumbinary planets on initially coplanar, circular orbits. Specifically, we find that the accuracy of our DNN never drops below 86%, even when tightly surrounding the boundary of instability, and is fast enough to be practical for on-the-fly calls during likelihood evaluations typical of modern Bayesian inference. Our binary classifier DNN is made publicly available at https://github.com/CoolWorlds/orbital-stability