# Predicting Instability Timescales in Closely-Packed Planetary Systems

Wednesday 4 July, 12:00

Many of the multi-planet systems discovered to date around other stars are maximally packed. This implies that N-body simulations with masses or orbital parameters too far from the actual values will destabilize on short timescales; thus, long-term integrations allow one to constrain the orbital architectures of many closely packed multi-planet systems. This technique has yielded insights into several important systems, e.g., HR 8799, Kepler 11, TRAPPIST-1, and our own solar system. However, a central challenge in such studies is the large computational cost of N-body simulations, which preclude a full survey of the high-dimensional parameter space of orbital architectures allowed by observations.

I will present our recent success in speeding up this dynamical characterization by a factor of 1 million using machine learning. In particular, we generated a million-CPU-hour dataset of N-body simulations of tightly packed systems, and trained a gradient-boosted decision tree algorithm (XGBoost) to predict stability over billion-orbit timescales. By optimizing dynamically informative features that we feed to the algorithm for each planetary system, we achieve a precision and recall of 90% on a holdout test set of N-body simulations from our dataset, as well as on N-body simulations of real Kepler systems.

I will discuss the wide discovery space that this opens up for characterizing multi-planet systems and for elucidating how orbital architectures evolve through time, just as the next generation of spaceborne exoplanet surveys launch this year.