Demolition Dash! Using Machine Learning to Model Stellar Collisions in the Galactic Center
Abstract
Direct collisions between stars in the cluster can alter the orbits and properties of stars, such as their mass. A subset of these collisions are mutually destructive events (MDEs), where the collision causes the complete disruption of both stars. However, the outcome of a collision depends on a number of parameters, including the mass of each star, age, impact parameter, and initial relative velocity, and models of dense stellar systems relying on simple collision prescriptions may not accurately represent the collision outcomes for such a large parameter space, while accurate collision prescriptions such as smoothed-particle hydrodynamic (SPH) simulations are computationally expensive for tens of thousands of collisions over the course of a simulation. We harness machine learning (ML) techniques to bridge high-precision hydrodynamics simulations of stellar collisions with our dynamical model. We train our ML model on a grid of ~27720 SPH simulations of collisions between main-sequence stars to accurately predict the collision-driven evolution of their masses. We run a semi-analytical model of a nuclear star cluster comprised of 1 M⊙ stars and compare the results of our model, the first of its kind, with previous simulations that relied on fitting formulae to predict collision outcomes. We find that our ML model's predictions align with previous studies of the galactic center and we examine the conditions that lead to MDEs. We find that MDEs are caused by a combination of high speed (~6000 km/s relative velocity) and nearly head-on (~0.2 impact parameter normalized to total radius) collisions of low mass <0.5M⊙ progenitor stars. The progenitors' low masses arise because they typically undergo 10 or more collisions over their lifetimes, gradually reducing their mass before they are ultimately destroyed.
Introduction
At the center of almost every galaxy lies a supermassive black hole (SMBH) and a dense cluster of stars surrounding it [1]. Interactions between these stars can alter their orbits and properties (such as mass, size, and age) [2]. There are two main types of interactions: relaxations and collisions. Relaxations refers to the cumulative effect of gravitational interactions with other stars in the cluster. Collisions refers to direct interactions between two stars. As a result of these interactions, a star can experience a Tidal Disruption Event (TDE) in which the star is torn apart by the tidal forces of the SMBH [3]. During this process, the matter being accreted to the SMBH releases large amounts of energy, producing an electromagnetic signature. Stars can also experience a Mutually Destructive Event (MDE) in which they have a high speed and direct collision with another star, destroying both. MDEs can potentially produce an electromagnetic transient which can be detected [4]. Stars can also become ejected from a cluster, merge with other stars, or destroy another star while remaining intact. With the rise of new sky surveys and better observatories, we expect to detect many more TDEs in the upcoming years.
Predicting the outcome of a collision from initial conditions can be difficult and computationally expensive.
To overcome the computational strain, we levearage machine learning (ML) to predict the outcome of any collision based
on a grid of smoothed particle hydrodynamic (SPH) results. Using deep learning, we can train a neural
network to identify the outcomes of star collisions based on a set of nearly 26000
simulations [5]. ML models typically falls into two main categories: classification and regression. Classification is used to
sort inputs into predetermined classes, such as apples and oranges, and regression is used to predict a continuous output. In our model,
we need classification to determine the type of collision that occurs (which could be an MDE, a merger, or a stripped star) and
regression to determine the change in mass of each star. Therefore, we apply a multi-task classification and regression neural network
to predict the outcome of stellar collisions.