Following Tracer Particles to Identify Filaments in Star-Forming Cores with HDBSCAN

Author: Nuray Ortaköse

Filaments appear on all scales of the cold, dense ISM. Observations show that star-forming cores are interlaced in complex networks of filaments. These structures are thought to funnel material onto forming cores through gravitationally driven accretion, yet the influence of filamentary dynamics on star formation remains uncertain. Observations struggle with line-of-sight confusion: structures that appear continuous in Position-Position-Velocity (PPV) space may not be connected in PPP (Clarke et al. 2018). While recent work has advanced filament identification and kinematic analysis (e.g. Chen et al. 2020), robust methods for identifying velocity-coherent filaments in simulations are still limited.

Simulations of colliding flows demonstrate that filaments can form self-consistently and act as accretion channels (e.g. Gómez & Vázquez-Semadeni 2014). In this work, I analyse tracer particles in a colliding-flow simulation from Weis et al. (2024). These tracer particles are massless, passive particles that follow the gas velocity field while recording physical quantities such as density and temperature.

To identify filamentary structures, I cluster tracer particles in a six-dimensional phase space (PPPVVV) using HDBSCAN. HDBSCAN is a non-parametric clustering algorithm that works well on data with various shapes and densities, while also identifying noise (i.e. not assigned to any cluster). This approach allows the identification of structures that are coherent in both position and velocity space. As a first validation, I compare filament spines extracted from the gas density field using DisPerSE (Sousbie 2011) with spines derived from clustered tracer particles (See Fig. 1). Near the center of mass (COM), both methods show strong agreement, demonstrating that tracer particles can reliably recover structures present in the underlying gas distribution.

In addition, I find that the scaling factor relating positional and velocity coordinates is critical for identifying physically meaningful structures through clustering (See Fig. 2). In particular, the dependence on α varies significantly between snapshots: some exhibit a broad range of high relative validity, indicating robust and well-defined clustering, while others show a narrow, isolated peak, suggesting that the clustering solution is highly sensitive to the choice of α. In some cases, clustering becomes extremely robust and largely independent of α. This diversity reflects differences in the dynamical state of the system rather than a simple temporal evolution.