How Large Eddy Simulations Are Cracking the Code of Cosmic Turbulence

Turbulence is everywhere in the universe. It churns in the convection zones of stars, whips through the solar wind, rages inside supernova explosions, and stirs the tenuous gas between galaxies. Yet simulating it faithfully has long been one of the most stubborn problems in computational astrophysics, because the range of length scales involved is so vast that no computer on Earth can track every swirl and eddy. A comprehensive review published in Living Reviews in Computational Astrophysics by Wolfram Schmidt-Brückner now takes stock of a powerful workaround: large eddy simulations, or LES, a technique borrowed from engineering and meteorology that deliberately computes only the largest, energy-carrying motions and models everything smaller with subgrid-scale physics.

The core obstacle is captured by a single number, the Reynolds number, which measures the ratio of non-linear inertial forces to viscous damping in a flow. When this number exceeds roughly a few thousand, flows become turbulent. Astrophysical flows routinely reach values that are astronomically larger; the turbulent convection zone of the Sun, for instance, is estimated to have a Reynolds number of order ten to the fourteenth. Direct numerical simulations that resolve every scale down to the microscopic dissipation length would require computing resources that scale with the Reynolds number cubed. For solar convection alone, the review notes, a single dynamical time would demand roughly ten to the forty-second floating point operations, far beyond even exascale machines.

Large eddy simulations sidestep this impossibility through a mathematical procedure called spatial filtering. A low-pass filter, defined by a convolution with a kernel such as a box, Gaussian, or sharp cut-off filter, separates the flow into smoothed, large-scale variables and fluctuating components on scales smaller than the filter length, which is usually identified with the grid spacing. Because the equations of fluid dynamics are non-linear, filtering them generates new terms: a subgrid-scale turbulence stress tensor that describes the momentum exchange between resolved and unresolved eddies, a turbulent pressure proportional to the subgrid-scale kinetic energy, and additional fluxes for energy and chemical species. The trace of the stress tensor defines the subgrid-scale turbulence energy, the kinetic energy hidden in eddies too small for the grid to see. Modelling these terms is known as the closure problem, since the filtered equations form an infinite hierarchy of moments that must be truncated.

Several families of closures dominate astrophysical practice. The simplest is the Smagorinsky model, which assumes a local balance between turbulence production and dissipation and yields an eddy viscosity proportional to the square of the grid scale times the local rate of strain. More sophisticated is the one-equation model, which solves a transport equation for the subgrid-scale turbulence energy itself, including production by shear, dissipation, turbulent diffusion, and pressure-dilatation effects. A third approach, structural modelling, reconstructs the unresolved stresses directly from gradients of the resolved velocity and magnetic fields, without tuneable coefficients or extra equations. For magnetohydrodynamics, the structural model also provides a closure for the subgrid-scale electromotive force, the term that governs how unresolved turbulent motions amplify magnetic fields through dynamo action.

A crucial subtlety is that most astrophysical simulation codes already dissipate energy through their numerical schemes. Finite-volume and finite-difference truncation errors behave like diffusion terms, producing an effective numerical viscosity that mimics turbulent viscosity on the grid scale. This observation underlies implicit large eddy simulation, or ILES, in which no explicit model is used and the numerics themselves play the role of the closure. The review explains why ILES often suffices: because inertial-range scaling is largely independent of the dissipation mechanism, statistics such as energy spectra can be reproduced as long as the dynamical range is adequate. However, the so-called bottleneck effect distorts spectra near the grid scale, and explicit subgrid-scale models add only a marginal effect for strongly diffusive second-order schemes. The picture changes for higher-order methods and mesh-free Lagrangian codes, which lack intrinsic diffusion and benefit far more from explicit modelling.

Validating these models without experiments or direct numerical simulations is a distinctive challenge of astrophysics. The standard tools are a priori tests, in which turbulence data from simulations are explicitly filtered and the correlations between modelled and true subgrid-scale terms are measured, and a posteriori comparisons of turbulence statistics between LES and ILES. Such analyses have calibrated the coefficients of the eddy-viscosity and gradient-diffusion closures, revealing, for example, that the turbulent Prandtl number is around ten rather than unity, and that purely linear eddy-diffusivity closures fail badly for magnetic stresses, matching the sign of the energy cascade only about half the time. Non-linear structural closures, by contrast, achieve correlations close to unity for the kinetic and magnetic stress tensors and the electromotive force, especially when a compressibility correction is included.

The payoff of this machinery appears across a remarkable range of cosmic phenomena. In thermonuclear supernovae, where a white dwarf detonates in a deflagration that no grid can resolve, the subgrid-scale turbulence energy sets the effective flame propagation speed, replacing the microscopic laminar burning velocity with a turbulent one. In galaxy simulations, the subgrid-scale turbulence energy can be fed by supernova feedback as an additional source term, and it allows a locally variable, turbulence-regulated star formation efficiency to be computed from the turbulent Mach number and virial parameter, reproducing observed star formation laws without assuming them. In cosmological simulations, filtered equations formulated in co-moving coordinates show that unresolved turbulent pressure can contribute a non-negligible fraction of the support against gravity in galaxy clusters, while shear-improved models cleanly separate genuine turbulence from gravity-driven bulk flows.

Perhaps the most striking recent success concerns magnetic field amplification in binary neutron star mergers. Kelvin-Helmholtz instabilities in the shear layer between the merging cores trigger a small-scale dynamo, and the growth rate of the magnetic field normally increases with resolution because the smallest resolved eddies dominate the amplification. By combining a relativistic generalization of the structural subgrid-scale model with high-order schemes, researchers achieved, for the first time, a numerically converged magnetic field amplification in merger simulations: the large eddy simulation reached the saturated field strength at half the resolution required without the model, a substantial saving in computational cost. The resulting magnetic energy spectrum even displays the expected Kazantsev scaling, and the post-merger field appears universal, insensitive to the initial magnetic configuration of the two stars.

Subgrid-scale modelling also matters for the chemical evolution of galaxies. In mesh-free codes without numerical diffusion, metals expelled by supernovae would never mix with their surroundings without an explicit turbulent diffusivity, and studies show that such mixing significantly reshapes the metallicity distributions of the circumgalactic and warm-hot intergalactic media. Looking forward, the review argues that the main utility of subgrid-scale models lies precisely in this treatment of complex sub-resolution physics, from turbulent burning and star formation to dynamo action and metal transport. While statistical properties of well-resolved turbulence remain largely insensitive to the choice of model, the coupling between resolved and unresolved scales is real and physically motivated closures capture it better than numerical truncation errors alone. As exascale computing opens new regimes, large eddy simulations are poised to remain an indispensable bridge between the eddies we can compute and the turbulent universe we observe.

Subject of Research: The methodology of large eddy simulations and subgrid-scale modelling for turbulent astrophysical flows

Article Title: Large eddy simulations in astrophysics

Article References: Schmidt-Brückner, W. (2025). Large eddy simulations in astrophysics. Living Reviews in Computational Astrophysics, 11(1), Article 2. https://doi.org/10.1007/s41115-025-00023-1

Image Credits: AI Generated

DOI: 10.1007/s41115-025-00023-1

Keywords: large eddy simulation, subgrid-scale modelling, astrophysical turbulence, magnetohydrodynamics, computational astrophysics, turbulent dynamo, type Ia supernovae, star formation, galaxy clusters, neutron star mergers, turbulent mixing, cosmological simulations