 |
Center for Nonlinear Studies |
Lagrangian (particle-based) methods are the dominant numerical tool for simulating complex boundaries and multiphase flows. However, while Eulerian frameworks have advanced significantly in turbulence closure modeling, such as Large Eddy Simulations (LES), Lagrangian frameworks lag behind. This presentation starts with the basics of Smoothed Particle Hydrodynamics (SPH), proceeds to address the structural reasons why turbulence modeling is inherently more cumbersome in Lagrangian settings, and finally proposes a Machine Learning (ML) solution. We first benchmark various SPH methods on fully-resolved, high-Reynolds-number 2D flows to establish a robust baseline. Subsequently, we introduce a data-driven approach for under-resolved settings, learning a turbulence closure specifically at the level of particle-particle interactions. Our results highlight the potential of ML-driven Lagrangian closures while identifying specific stability challenges inherent to this hybrid approach.