The problem begins with the sheer range of scales involved. In the final orbit before two neutron stars merge, the stellar cores slam into one another at a substantial fraction of the speed of light, generating a shear layer roughly a kilometer wide that becomes Kelvin-Helmholtz unstable. This instability, the same mechanism that shapes wind-blown clouds on Earth, shreds the interface between the stars into vortices that fragment into ever smaller eddies, producing a turbulent cascade that spans from kilometer scales down to about a nanometer, where viscosity finally converts kinetic energy into heat. The Reynolds number of this flow, a measure of the ratio of inertial to viscous forces, is a staggering ten to the power of sixteen. Simulating every eddy directly, the approach known as direct numerical simulation, would require computational resources that scale as the Reynolds number cubed, making it utterly impossible for the foreseeable future.

Turbulence is not confined to the moment of contact. Once the stars have merged, the remnant, whether a massive neutron star or a newly formed black hole, is typically encircled by a hot, dense accretion disk. There, the magnetorotational instability stirs the plasma, redistributing angular momentum and governing how matter spirals inward or is flung outward. The way turbulence transports angular momentum determines whether the remnant neutron star collapses promptly to a black hole or survives as a long-lived object, and it controls the mass ejection that powers the kilonova flashes and the nucleosynthesis of r-process elements. It may also amplify magnetic fields to magnetar levels, potentially launching the relativistic jets that produce short gamma-ray bursts. In short, nearly every observable signature of a neutron star merger is touched by turbulence somewhere along the way.

The mathematical machinery for handling unresolved turbulence has a long history in Newtonian fluid dynamics. The classic approach, Reynolds averaging, splits the flow into a mean component and fluctuations, yielding equations for the mean motion that contain an extra term, the Reynolds stress, which encapsulates the momentum carried by the unresolved eddies. A more practical alternative for simulations is large-eddy simulation, or LES, in which the equations are filtered over a length scale comparable to the numerical grid. The filtered equations resemble the original ones but include subgrid-scale stresses that must be modeled. The central difficulty, known as the closure problem, is that these stresses depend on information about the small scales that the simulation does not compute, so modelers must supply approximate relations, or closures, that capture the net effect of the missing physics using only the resolved quantities.

Extending this framework to general relativity introduces subtleties that have no Newtonian counterpart. Radice and Hawke review how averaging or filtering the equations of relativistic hydrodynamics produces effective stresses even when the underlying fluid is ideal, and how the nonlinear structure of the fluxes demands additional closure relations, including one for turbulent mass diffusion. More troubling still is the question of covariance: the averaging operations used in practice are tied to a particular slicing of spacetime, which breaks the four-dimensional symmetry of Einstein’s theory. Recent work has explored building the averaging procedure around a physical observer rather than a coordinate slice, showing that the coarse-grained equations then take the form of a non-ideal relativistic fluid, complete with bulk viscosity, shear stresses, and heat transport terms that arise purely from the turbulence. Even the equation of state, the relation linking pressure, density, and energy, is modified by averaging, since fluctuations in density generate corrections that behave like an additional pressure.

In practice, most published neutron star merger simulations to date have used the simplest possible strategy: implicit large-eddy simulation, which sets the subgrid stresses to zero and relies on the intrinsic numerical dissipation of shock-capturing schemes to mimic the effect of unresolved turbulence. This approach has been remarkably successful in other fields, but the review is blunt about its limitations in this context. The modified equation analysis shows that numerical dissipation can indeed act like an effective viscosity, but implicit methods require a significant fraction of the inertial range to be resolved before results converge, and no neutron star merger simulation has yet been demonstrated to be in that regime. The alternative is explicit modeling. Radice’s own general-relativistic large-eddy simulations employ a relativistic version of the Smagorinsky closure, in which the turbulent viscosity is estimated from a mixing length set by the local scale of the flow and the speed of sound. A third family of methods, gradient or approximate-deconvolution models, reconstructs the effect of the filter algebraically and has the advantage of introducing no tunable parameters beyond the filter width itself.

The payoff of these techniques is already visible in the study of magnetic field amplification. Early Newtonian simulations suggested that the Kelvin-Helmholtz instability could amplify even weak seed fields to magnetar strengths of around ten to the fifteenth gauss, but general-relativistic calculations initially failed to reproduce this, simply because their grids were too coarse. Later, extraordinarily high-resolution simulations by Kenta Kiuchi and collaborators showed that the saturated field strength kept climbing with resolution, with no sign of convergence, precisely because the magnetic back-reaction only halts the cascade at centimeter scales, far below anything a global simulation can resolve. When subgrid models were introduced, the picture changed dramatically. Gradient-model simulations by Ricard Aguilera-Miret, Carlos Palenzuela, and colleagues achieved converged results, confirming that weak fields are indeed amplified to ten to the sixteenth gauss and that the statistical properties of the resulting turbulence are remarkably insensitive to the unknown initial magnetic configuration inside the stars, a reassuring result for predictive modeling.

The same simulations revealed tantalizing evidence of an inverse cascade, in which the characteristic scale of the magnetic field grows from roughly half a kilometer immediately after merger to several kilometers a hundred milliseconds later, as turbulent resistivity rearranges field lines into larger structures. Meanwhile, measurements of the effective viscosity generated by magnetic stresses suggest it remains modest in the dense core of the remnant, implying that turbulence is unlikely to distort the post-merger gravitational-wave signal enough to compromise plans to probe the equation of state of nuclear matter with next-generation detectors such as the Einstein Telescope and Cosmic Explorer. On the other hand, turbulence and dynamo action are expected to leave a significant imprint on the long-term evolution of the remnant, its mass ejection, and its multi-messenger emission, from kilonova light curves to the engines of short gamma-ray bursts.

Much remains to be done. The review highlights open questions about whether angular momentum transport accelerates or delays the collapse of the remnant neutron star, about the topology of the amplified magnetic fields and whether tangled configurations better explain the energetics of gamma-ray bursts, and about the formidable challenge of uncertainty quantification in a parameter space already crowded with uncertain inputs. Validation is particularly thorny: unlike wind tunnels, neutron stars offer no laboratory tests, so models must be calibrated against resolved simulations whose own fidelity is uncertain, and tuned to observables, such as gravitational waves and neutrino signals, that differ from the statistical quantities conventionally used in closure validation. The authors anticipate rapid progress on three fronts: simulations that combine sophisticated microphysics, magnetohydrodynamics, and large-eddy closures; improved phenomenological subgrid models tested in local calculations; and data-driven, machine-learned closures that learn the missing physics directly from high-resolution data. As gravitational-wave astronomy enters its next generation, taming turbulence may prove the key to turning collisions of dead stars into precision measurements of matter at its densest.

Subject of Research: Turbulence modelling in general-relativistic simulations of binary neutron star mergers

Article Title: Turbulence modelling in neutron star merger simulations

Article References: Radice, D., & Hawke, I. (2024). Turbulence modelling in neutron star merger simulations. Living Reviews in Computational Astrophysics, 10(1), Article 1. https://doi.org/10.1007/s41115-023-00019-9

Image Credits: AI Generated

DOI: 10.1007/s41115-023-00019-9

Keywords: neutron star mergers, turbulence, large-eddy simulation, gravitational waves, magnetohydrodynamics, Kelvin-Helmholtz instability, magnetic field amplification, subgrid models, numerical relativity, accretion disks, r-process nucleosynthesis, equation of state

The scale of the computational challenge is staggering. Temperature contrasts in astrophysical multiphase systems span roughly ten orders of magnitude, and the cold gas, while dense, can be scattered in tiny structures across enormous volumes. Observations suggest cold clouds in the circumgalactic medium may be as small as tens of parsecs. To fully resolve such structures across a Milky Way-sized halo would require more than a quadrillion resolution elements, a number so vast that no supercomputer on Earth comes close. As a result, cold gas quantities in typical cosmological simulations remain unconverged, and researchers have had to resort to a hierarchy of idealized experiments, from turbulent mixing layers to cloud-crushing simulations, each isolating one piece of the physics.

One of the central results the review highlights concerns the survival of cold clouds blasted by hot winds, the classic ‘cloud crushing’ problem. In the absence of cooling, a cold cloud embedded in a supersonic wind is shredded by Kelvin-Helmholtz and Rayleigh-Taylor instabilities on a characteristic crushing time. Worse, the drag time needed to accelerate the cloud is far longer than the destruction time, giving rise to the long-standing ‘entrainment problem’: how can cold gas be flung to hundreds or even thousands of kilometers per second in galactic winds if it should be destroyed before it can accelerate? The answer, it turns out, lies in radiative cooling. When mixed gas at the interface cools faster than the cloud is disrupted, hot gas condenses onto the cold phase, allowing clouds not only to survive but to grow in mass as they are carried along.

This insight has been crystallized into a simple survival criterion: clouds endure if the cooling time of the mixed gas is shorter than their destruction time, which can be recast as a minimum cloud size of roughly a few parsecs under typical wind conditions. Simulations show that surviving clouds grow continuously through cooling-driven mass transfer in their turbulent wakes, and that this same mixing efficiently transfers momentum, accelerating even dense molecular clouds entrained in galactic outflows. Magnetic fields, once heralded as a potential savior of the entrainment problem, help but are not sufficient on their own for the high density contrasts typical of real astrophysical clouds; combined with cooling, however, they shift the survival threshold by orders of magnitude.

The review also delves into thermal instability, the classic mechanism by which a hot medium can spontaneously fragment into a cold, clumpy phase. When radiative cooling increases as temperature drops, small density perturbations run away into dense clumps. In stratified halo atmospheres, precipitation occurs when the ratio of cooling time to free-fall time falls below a critical value of order ten, a criterion modified by turbulence, halo rotation, magnetic fields and cosmic rays. A related and still contentious question is whether cooling clouds ‘shatter’ into a characteristic scale of tiny fragments or instead undergo a violent pulsation dubbed ‘splattering’ before fragmenting, with recent three-dimensional simulations tending to favor the latter picture.

Scaling up, the review surveys supernova-driven bubbles, stratified ‘tall box’ simulations of galaxy disks, and fully global models of dwarf and Milky Way-mass galaxies. A consistent picture emerges: most outflowing mass travels in the warm phase at around 10,000 Kelvin, while most of the energy is carried by the hot, million-degree gas. Hot gas mass loading factors hover near 0.1 across a wide range of star formation rates, and warm outflows in massive galaxies tend to fall back as fountain flows rather than escaping. Including cosmic rays transforms these results, converting fountains into steady, cooler, denser winds that can double the outflowing mass and substantially reshape the circumgalactic medium.

At the largest scales, the review examines how simulations handle the circumgalactic and intracluster media. In cluster cores, jets from supermassive black holes stir turbulence that triggers local thermal instability, producing ‘chaotic cold accretion’ in which cold filaments rain onto the central galaxy and feed the black hole in a self-regulating cycle. In cosmological zoom-in simulations, a recent revolution has come from ‘super-Lagrangian’ refinement schemes that boost resolution specifically in the halo, reaching below 100 parsecs in the circumgalactic medium. These enhanced-resolution models consistently show more and smaller cool clouds, higher covering fractions of cool gas, and non-converged cloud mass functions, confirming that cold gas structure in halo simulations is far from fully resolved.

What emerges most clearly from this comprehensive synthesis is that the diverse simulation approaches, from idealized mixing layers to full cosmological models, are not competitors but complementary layers of a single framework. Small-scale experiments provide the physical intuition, survival criteria and subgrid prescriptions that large-scale simulations need; large-scale simulations in turn supply the realistic boundary conditions, pressures and turbulence levels under which the small-scale physics operates. The ultimate arbiter, the authors stress, is observation, and connecting simulations to real spectra, emission maps and absorption measurements through radiative transfer remains one of the field’s most demanding tasks.

The challenges ahead are formidable: achieving numerical convergence in multiphase diagnostics, capturing the interplay of magnetic fields, conduction, viscosity and cosmic rays, and resolving the critical scales that govern whether cold gas survives, grows or shatters. But the trajectory is clear. GPU-accelerated codes, adaptive refinement targeted at cooling lengths, and a maturing theoretical framework are converging on a unified picture of the multiphase Universe, one simulation at a time.

Subject of Research: Numerical simulations of multiphase gas dynamics in the interstellar, circumgalactic and intracluster media

Article Title: Simulations of multi-phase gas in and around galaxies

Article References: Gronke, M., & Schneider, E. E. (2026). Simulations of multi-phase gas in and around galaxies. Living Reviews in Computational Astrophysics, 12(1), Article 2. https://doi.org/10.1007/s41115-026-00025-7

Image Credits: AI Generated

DOI: 10.1007/s41115-026-00025-7

Keywords: multiphase gas, galaxy simulations, circumgalactic medium, thermal instability, galactic winds, cloud crushing, turbulent mixing layers, computational astrophysics, radiative cooling, cosmic rays, supernova feedback, Kelvin-Helmholtz instability