Particle Clustering in Turbulence: Prediction of Spatial and Statistical Properties with Deep Learning

Yan-Mong Chan, Natascha Manger, Yin Li, Chao‐Chin Yang, Zhaohuan Zhu, Philip J. Armitage, Shirley Ho

TL;DR

Researchers have successfully trained a 3D U-Net deep learning model to predict how dust particles cluster within turbulent gas—a key process in the birth of planets. By mapping fluid velocity and density directly to particle distributions, the model captures complex filamentary structures and recovers critical statistical properties with less than 10% error, offering a massive speedup over traditional numerical simulations.

What Problem Does This Paper Address?

In the swirling disks of gas and dust around young stars (protoplanetary disks), the journey from microscopic dust to planetesimals is driven by turbulence. Turbulence causes particles to cluster, which in turn dictates how often they collide and grow.

While we have powerful tools like Direct Numerical Simulations (DNS) to model these interactions, they are computationally "expensive." Simulating a wide range of particle sizes across vast spatial scales is often prohibitive. Analytic models exist but struggle to capture the highly nonlinear, "filamentary" nature of particle clustering. The authors sought to determine if deep learning could bridge this gap by learning the complex mapping between a turbulent fluid and the particles suspended within it.

Methods

The team used the Athena++ hydrodynamics code to generate a high-fidelity dataset of particles in a "periodic box" of isotropic turbulence. They focused on the Epstein drag regime, which is physically relevant for micron- to millimeter-sized dust in the low-density environment of a protoplanetary disk.

The core of their approach is a 3D U-Net, a convolutional neural network architecture famous for its ability to handle volumetric data.

  • Input: 3D grids of gas density and gas velocity.
  • Output: 3D grids of particle density and particle velocity.

A key technical innovation was training the model on relative velocity ($v_{rel} = v_{particle} - v_{gas}$) rather than absolute particle velocity. This "UNET-R" approach prevents the bulk gas flow from masking the subtle, physically interesting differences in how particles move compared to the gas.

Datasets

The researchers generated their own training data using Lagrangian particle simulations:

  • Size: 252 snapshots of $256^3$ resolution grids.
  • Augmentation: To make the most of the data, they used random $128^3$ crops, rotations, and parity flips during training.
  • Physics: The particles were assigned a Stokes number ($St$)—a dimensionless measure of how well a particle follows the fluid flow—of approximately 1, where clustering effects are most pronounced.

Key Insights and Ideas

  • The "Relative" Advantage: Training on relative velocity was significantly more effective than training on absolute velocity. It forced the network to learn the physics of the coupling between gas and dust rather than just learning to copy the background gas movement.
  • Statistical Fidelity from Spatial Training: Even though the loss function only looked at voxel-by-voxel differences (Mean Squared Error), the model "accidentally" learned to respect complex statistical distributions. It accurately reproduced the Radial Distribution Function (RDF), which measures the probability of finding two particles at a certain distance from each other.
  • Smoothing vs. Sharpness: Like many CNN-based generators, the model tended to produce slightly "smoother" results than the ground truth. While it captured the locations of filaments and voids perfectly, it slightly underpredicted the extreme density peaks.
  • Generalization Limits: The model generalized well to different "driving scales" of turbulence. However, it struggled when the fundamental physics changed (e.g., changing the particle size/friction time). This suggests the model learns the specific mapping for a physical regime rather than a universal law of motion.

Key Results

  • Accuracy: The model predicted density and velocity statistics with errors typically under 10%.
  • Qualitative Success: Visual inspections of the 3D slices show that the U-Net captures the "web-like" filamentary structures of dust clustering that are notoriously difficult to model analytically.
  • Performance: Once trained, the U-Net can produce a 3D particle field almost instantaneously, compared to the hours or days required for a full Lagrangian simulation.

Limitations and Future Work

The authors acknowledge a few hurdles before this can be used in a production-level astrophysical pipeline:

  1. Parameter Sensitivity: Currently, a new model must be trained for every particle size (Stokes number). Future work aims to include the Stokes number as an input parameter so one model can handle a full distribution of dust sizes.
  2. One-Way Coupling: This study assumed the gas moves the dust, but the dust doesn't move the gas ("one-way coupling"). In dense regions, the "backreaction" of dust on gas is vital and would require a more complex model.
  3. Resolution: While $256^3$ is a standard benchmark, real-world applications may require even higher resolutions to capture the smallest scales of collision physics.

By demonstrating that deep learning can "see" the relationship between gas turbulence and dust clustering, this paper opens the door to much larger, more complex simulations of how planets begin their lives.

Summary updated

Abstract

Abstract We investigate the utility of deep learning for modeling the clustering of particles that are aerodynamically coupled to turbulent fluids. Using a Lagrangian particle module within the Athena++ hydrodynamics code, we simulate the dynamics of particles in the Epstein drag regime within a periodic domain of isotropic forced hydrodynamic turbulence. This setup is an idealized model relevant to the collisional growth of micron- to millimeter-sized dust particles in early-stage planet formation. The simulation data are used to train a U-Net deep-learning model to predict gridded three-dimensional representations of the particle density and velocity fields, given as input the corresponding fluid fields. The trained model qualitatively captures the filamentary structure of clustered particles in a highly nonlinear regime. We assess model fidelity by calculating metrics of the density field (the radial distribution function) and of the velocity field (the relative velocity and the relative radial velocity between particles). Although trained only on the spatial fields, the model predicts these statistical quantities with errors that are typically <10%. Our results suggest that, given appropriately expanded training data, deep learning could complement direct numerical simulations in predicting particle clustering within turbulent flows.
BibTeX
@article{chan2023particle,
  author = {Yan-Mong Chan and Natascha Manger and Yin Li and Chao‐Chin Yang and Zhaohuan Zhu and Philip J. Armitage and Shirley Ho},
  title = {{Particle Clustering in Turbulence: Prediction of Spatial and Statistical Properties with Deep Learning}},
  journal = {The Astrophysical Journal},
  year = {2023},
  doi = {10.3847/1538-4357/ad088c},
}
Figure