ML in Astrophysical Turbulence I: Predicting Prestellar Cores in Magnetized Molecular Clouds using eXtreme Gradient Boosting

TL;DR

This paper introduces a machine learning framework using XGBoost to predict the future evolution of prestellar cores in turbulent magnetized molecular clouds, achieving high accuracy and offering a new tool for star formation modeling.

Contribution

The study presents the first application of supervised machine learning to forecast star-forming core evolution in MHD turbulence, demonstrating high predictive accuracy with phase-space data.

Findings

Achieved R^2 > 0.99 in predicting core evolution

Successfully distinguished collapsing cores from transient fluctuations

Provided a computationally efficient alternative to sink-particle methods

Abstract

Giant Molecular Clouds (GMCs) are dominated by supersonic turbulence, creating a complex network of shocks and filaments that regulate star formation. While the global inefficiency of star formation is well-observed, predicting exactly which gas parcels within a turbulent cloud will collapse to form stars remains a challenge. In this work, we present a supervised machine learning framework to forecast the Lagrangian history of prestellar cores in magnetohydrodynamic (MHD) turbulence. We utilize Extreme Gradient Boosting (XGBoost) to train a regression model on the trajectories of $\sim 2.1$ million tracer particles evolved within a self-gravitating, turbulent MHD simulation. By mapping the instantaneous phase-space state (position, velocity, and density) of gas parcels to their future coordinates, our model successfully predicts the 3D evolution of star-forming cores over a horizon of $\sim 0.45$ Myr ($0.25~t_{\rm ff}$). We achieve a global coefficient of determination of $R^2 > 0.99$ and demonstrate that the model captures the non-linear convergent flows characteristic of gravitational collapse. Crucially, we show that local phase-space information alone is sufficient to distinguish between transient density fluctuations and bound collapsing cores. This data-driven approach offers a computationally efficient alternative to traditional sink-particle algorithms and provides a pathway for developing high-fidelity subgrid models for galaxy-scale simulations.