Fact-checked by Grok 2 weeks ago

Large eddy simulation

Large eddy simulation (LES) is a computational fluid dynamics technique for simulating turbulent flows, in which the large-scale, energy-containing eddies are directly resolved while the effects of smaller subgrid-scale eddies are modeled to account for their influence on the resolved scales.^[1] This approach applies a spatial filter to the Navier-Stokes equations, separating the flow into resolved and unresolved components, and requires closure models to approximate the subgrid-scale stresses that arise from the filtering operation.^[2] LES occupies an intermediate position between direct numerical simulation (DNS), which resolves all scales at high computational cost, and Reynolds-averaged Navier-Stokes (RANS) methods, which average out turbulence and model all unsteady effects, offering a balance of accuracy for unsteady, three-dimensional flows and reduced expense compared to DNS.^[1] The origins of LES trace back to 1963, when Joseph Smagorinsky introduced the concept in the context of numerical weather prediction, proposing an eddy viscosity model to parameterize subgrid-scale dissipation in simulations of atmospheric circulation using the primitive equations.^[3] Early developments focused on simple "building-block" flows like homogeneous isotropic turbulence and plane channel flows, with foundational filtering theory formalized by Leonard in 1974.^[2] Subsequent advancements included dynamic subgrid-scale models by Germano et al. in 1991, which adapt model coefficients based on resolved flow information to improve accuracy in non-equilibrium conditions.^[1] Common subgrid-scale models include the Smagorinsky eddy viscosity approach, with a constant typically ranging from 0.18 to 0.23, and scale-similarity or mixed models that better capture backscatter and intermittency.^[1] LES has become widely applied in engineering and geophysical contexts, such as aerodynamics, combustion, atmospheric boundary layers, and ocean currents, due to its ability to predict unsteady phenomena like vortex shedding and scalar mixing that RANS often underresolves.^[1] Key challenges include accurate modeling near walls, handling compressible flows and shocks, and managing high computational demands for complex geometries, often requiring advanced numerical schemes for stability and grid resolution on the order of the filter width.^[2] Ongoing research emphasizes regularization techniques, such as spectral viscosity or Leray-alpha models, to ensure mathematical well-posedness and physical realism in the filtered equations. Recent advancements as of 2025 include data-driven approaches using machine learning for subgrid-scale modeling and improved wall-modeling techniques for high-Reynolds-number flows.^[2]^[4]

Overview

Definition and Principles

Large eddy simulation (LES) is a hybrid numerical method in computational fluid dynamics used to simulate turbulent flows by directly resolving the large-scale eddies while modeling the effects of the unresolved small-scale eddies on the resolved motion.^[5] This approach bridges the gap between the full resolution required by direct numerical simulation (DNS) and the statistical averaging of Reynolds-averaged Navier-Stokes (RANS) methods, enabling detailed predictions of unsteady turbulent structures in complex geometries.^[5] In turbulent flows, energy production predominantly occurs at large scales, where eddies are anisotropic, intermittent, and strongly influenced by the mean flow, boundaries, and external forcing. These large eddies carry the majority of the kinetic energy and momentum transport, whereas smaller eddies, particularly those in the inertial and dissipative ranges, become increasingly isotropic and focus on cascading energy toward viscous dissipation without significant production. LES capitalizes on this scale separation by applying a spatial filter to the governing equations, distinguishing grid-scale (resolved) large eddies from subgrid-scale (unresolved) small eddies that are too fine to capture explicitly on practical meshes.^[5] The filtering operation decomposes the velocity field into resolved and subgrid components, with the subgrid-scale effects appearing as additional stress terms in the filtered equations.^[3] These subgrid-scale (SGS) stresses quantify the nonlinear interaction and momentum exchange between resolved and unresolved scales, necessitating closure models to approximate their influence without resolving all small-scale details.^[5] By targeting high-Reynolds-number flows where DNS demands grid resolutions scaling as Re^9/4 (prohibitively expensive for practical applications), LES achieves computational efficiency through coarser grids that resolve only the energy-containing large scales, typically reducing costs by orders of magnitude compared to DNS while retaining key unsteady features.^[5] The resulting filtered Navier-Stokes equations thus provide a framework for simulating the dynamics of large eddies accurately.^[5]

Historical Development

The concept of large eddy simulation (LES) emerged in the 1960s with Joseph Smagorinsky's development of the first subgrid-scale (SGS) closure model, an eddy-viscosity formulation inspired by Ludwig Prandtl's mixing-length theory, initially applied to simulate global atmospheric circulations using the primitive equations.^[3] This model parameterized the effects of unresolved small-scale eddies through a constant coefficient tied to the filter width, marking the foundational approach to balancing computational feasibility with turbulence representation in three-dimensional simulations.^[3] Advancements in the 1970s built on this foundation, with James W. Deardorff conducting the pioneering three-dimensional LES of turbulent channel flow at large Reynolds numbers, explicitly resolving large-scale motions while employing Smagorinsky's SGS model to capture subgrid effects. Deardorff's work demonstrated the practical viability of LES for non-atmospheric flows and introduced early techniques for adjusting model coefficients based on flow conditions, precursors to more sophisticated dynamic procedures. These simulations highlighted LES's potential for studying energy transfer across scales in homogeneous and shear-driven turbulence.^[6] The 1980s and 1990s saw rapid methodological evolution, culminating in the dynamic Smagorinsky model proposed by Mario Germano, Umberto Piomelli, and Parviz Moin in 1991, which dynamically computed scale-dependent coefficients using the Germano identity to relate resolved stresses at different filter levels, eliminating the need for a priori tuning. This innovation improved model adaptability for inhomogeneous flows and became a cornerstone of modern LES. A pivotal milestone during this period was the first wall-resolved LES of high-Reynolds-number turbulent boundary layers in the early 1990s, as demonstrated by Piomelli and colleagues, which accurately captured near-wall coherent structures and velocity profiles without wall functions. From the 2000s onward, LES integrated with advanced numerical schemes, including implicit LES (ILES), where numerical truncation errors provide the necessary SGS dissipation without explicit models, as formalized in monotone-integrated approaches by Grinstein and colleagues.^[7] The 2010s witnessed substantial growth in multiphysics LES, particularly for combustion and multiphase flows, enabled by refined SGS closures and high-performance computing to handle reactive and interfacial dynamics.^[8] Emerging machine learning-based SGS models, leveraging neural networks trained on high-fidelity data, further enhanced predictive capabilities for complex turbulence interactions.^[9] Post-2020 developments emphasize high-fidelity LES on exascale platforms, particularly for climate and environmental modeling, enabling finer resolutions to better resolve turbulent processes in clouds and boundary layers. These trends underscore LES's transition from theoretical tool to essential method for multiscale environmental simulations.

Mathematical Foundations

Filter Definition and Properties

In large eddy simulation (LES), the filtering operation is a low-pass mathematical process that separates the large-scale, resolved motions from the small-scale, subgrid-scale (SGS) fluctuations in turbulent flows. The filtered velocity field \bar{\mathbf{u}}(\mathbf{x}) is defined as the convolution of the unfiltered velocity \mathbf{u}(\mathbf{x}) with a filter kernel G(\mathbf{x} - \boldsymbol{\xi}, \Delta), where \Delta denotes the filter width:

\bar{\mathbf{u}}(\mathbf{x}) = \int G(\mathbf{x} - \boldsymbol{\xi}, \Delta) \mathbf{u}(\boldsymbol{\xi}) \, d\boldsymbol{\xi},

with the kernel satisfying normalization \int G(\mathbf{x}, \Delta) \, d\mathbf{x} = 1 and typically being positive and even for physical consistency.^[10] Common filter types include the box (top-hat) filter, Gaussian filter, and spectral (sharp cutoff) filter, each with distinct transfer functions in Fourier space that determine the scale separation. The box filter kernel is G(\mathbf{x}, \Delta) = \Delta^{-3} for |\mathbf{x}| < \Delta/2 and zero otherwise, yielding transfer function \hat{G}(\mathbf{k}) = \mathrm{sinc}^3(k \Delta / 2), where \mathrm{sinc}(\theta) = \sin(\theta)/\theta. The Gaussian filter uses G(\mathbf{x}, \Delta) = (6/\pi \Delta^2)^{3/2} \exp(-6 |\mathbf{x}|^2 / \Delta^2), with transfer function \hat{G}(\mathbf{k}) = \exp(-k^2 \Delta^2 / 24). The spectral filter applies an abrupt cutoff at wavenumber k_c = \pi / \Delta, so \hat{G}(\mathbf{k}) = 1 for k < k_c and 0 otherwise. These filters vary in their localization: the box filter is compact in physical space but oscillatory in spectral space, the Gaussian is smooth in both, and the spectral is ideal for scale separation but non-local in physical space. Filters in LES possess key mathematical properties that ensure robust scale separation and numerical stability. Linearity holds for all convolution-based filters, allowing superposition: \overline{a \mathbf{u} + b \mathbf{v}} = a \bar{\mathbf{u}} + b \bar{\mathbf{v}}. Commutativity with spatial differentiation is exact for translationally invariant filters in continuous space, such as the box and spectral filters. Idempotence, where double filtering equals single filtering (\overline{\bar{\mathbf{u}}} = \bar{\mathbf{u}}), is exact for the sharp spectral filter but only approximate for others like the Gaussian. Realizability requires a positive-definite kernel to preserve physical quantities like non-negative densities and avoid unphysical oscillations.^[2] Applying the filter to nonlinear terms, such as products in the Navier-Stokes equations, introduces unresolved contributions. The standard decomposition, known as the Leonard decomposition, breaks down the subgrid-scale (SGS) stress tensor \tau_{ij} = \overline{u_i u_j} - \bar{u}_i \bar{u}_j into three parts: the Leonard stress L_{ij} = \widetilde{\bar{u}_i \bar{u}_j} - \tilde{u}_i \tilde{u}_j, the cross stress C_{ij} = \overline{\tilde{u}_i u_j'} + \overline{u_i' \tilde{u}_j}, and the Reynolds stress R_{ij} = \overline{u_i' u_j'}, where \tilde{\cdot} denotes a test filter (typically wider than the primary filter), and primes denote deviations from the test-filtered field. This decomposition identifies the resolved-resolved interactions (Leonard term, computable from resolved scales) separately from cross and sub-test contributions that require modeling.^[10] Filter width \Delta is selected to lie between the integral scale and the Kolmogorov scale, typically 2–4 times the grid spacing \Delta_x to ensure that resolved scales dominate the energy-containing motions while keeping computational costs manageable; smaller ratios risk inadequate resolution, while larger ones increase modeling demands.^[11]

Filtered Governing Equations

Large eddy simulation (LES) begins with the application of a low-pass spatial filter to the governing equations of fluid motion, separating the resolved large-scale structures from the unresolved subgrid-scale (SGS) motions. For incompressible flows, the starting point is the unfiltered Navier-Stokes equations and continuity equation:

\frac{\partial u_i}{\partial t} + u_j \frac{\partial u_i}{\partial x_j} = -\frac{1}{\rho} \frac{\partial p}{\partial x_i} + \nu \frac{\partial^2 u_i}{\partial x_j \partial x_j},

\frac{\partial u_i}{\partial x_i} = 0,

where u_i is the velocity component, p is the pressure, \rho is the constant density, and \nu is the kinematic viscosity. Applying the filter operation, denoted by an overbar (e.g., \bar{f} = \int G(\mathbf{x} - \boldsymbol{\xi}) f(\boldsymbol{\xi}) d\boldsymbol{\xi}, where G is the filter kernel), to these equations yields the filtered continuity equation

\frac{\partial \bar{u}_i}{\partial x_i} = 0,

which holds because the divergence-free condition is preserved under filtering for incompressible flows, assuming the filter commutes with spatial derivatives. The filtered momentum equation becomes

\frac{\partial \bar{u}_i}{\partial t} + \frac{\partial (\bar{u}_i \bar{u}_j)}{\partial x_j} = -\frac{1}{\rho} \frac{\partial \bar{p}}{\partial x_i} + \nu \frac{\partial^2 \bar{u}_i}{\partial x_j \partial x_j} - \frac{\partial \tau_{ij}}{\partial x_j},

where the nonlinear convection term \overline{u_j \partial_j u_i} leads to the SGS stress tensor \tau_{ij} = \overline{u_i u_j} - \bar{u}_i \bar{u}_j after applying the product rule and the incompressibility constraint, which introduces the unresolved contribution from the small scales. This SGS term represents the effect of the filtered-out scales on the resolved motion and must be modeled in LES. For compressible flows, where density variations are significant, the standard spatial filtering leads to additional unclosed terms in the filtered equations due to the variable density. To address this, Favre filtering—a density-weighted averaging—is employed, defined as \tilde{f} = \overline{\rho f} / \bar{\rho}, which simplifies the form of the equations similar to the incompressible case. The unfiltered compressible Navier-Stokes equations include continuity, momentum, and energy equations, but the filtering process introduces SGS terms for mass, momentum, and energy fluxes. The Favre-filtered continuity equation is

\frac{\partial \bar{\rho}}{\partial t} + \frac{\partial (\bar{\rho} \tilde{u}_j)}{\partial x_j} = 0,

reflecting the filtered density and Favre-filtered velocity. The Favre-filtered momentum equation takes the form

\frac{\partial (\bar{\rho} \tilde{u}_i)}{\partial t} + \frac{\partial (\bar{\rho} \tilde{u}_i \tilde{u}_j)}{\partial x_j} = -\frac{\partial \bar{p}}{\partial x_i} + \frac{\partial \bar{\sigma}_{ij}}{\partial x_j} - \frac{\partial (\bar{\rho} \tau_{ij})}{\partial x_j},

where \bar{\sigma}_{ij} is the filtered viscous stress tensor, and the SGS stress \tau_{ij} = \widetilde{u_i u_j} - \tilde{u}_i \tilde{u}_j appears in Favre-averaged form, along with additional SGS terms arising from density-velocity correlations. The energy equation similarly involves filtered total energy with SGS heat flux and other unclosed terms. This formulation ensures that the resolved scales are governed by equations analogous to the incompressible case, but with density-weighting to account for variable-density effects prevalent in high-speed or reacting flows.

Filtered Kinetic Energy Equation

The evolution equation for the resolved kinetic energy in large eddy simulation (LES) is derived by taking the inner product of the filtered momentum equation with the filtered velocity field \bar{\mathbf{u}}, assuming an incompressible flow with constant density \rho = 1. This yields the balance for the resolved turbulent kinetic energy K_r = \frac{1}{2} \bar{u}_i \bar{u}_i, highlighting the production, dissipation, and transfer mechanisms across scales. Starting from the filtered incompressible Navier-Stokes equations, the derivation proceeds as follows. Contracting the filtered momentum equation \frac{\partial \bar{u}_i}{\partial t} + \frac{\partial (\bar{u}_i \bar{u}_j)}{\partial x_j} = -\frac{\partial \bar{p}}{\partial x_i} + \nu \frac{\partial^2 \bar{u}_i}{\partial x_j \partial x_j} - \frac{\partial \tau_{ij}}{\partial x_j} with \bar{u}_i and summing over i gives:

\frac{\partial}{\partial t} \left( \frac{1}{2} \bar{u}_k \bar{u}_k \right) + \frac{\partial}{\partial x_j} \left( \frac{1}{2} \bar{u}_k \bar{u}_k \bar{u}_j \right) = -\bar{u}_i \frac{\partial \bar{p}}{\partial x_i} + \nu \bar{u}_i \frac{\partial^2 \bar{u}_i}{\partial x_j \partial x_j} - \bar{u}_i \frac{\partial \tau_{ij}}{\partial x_j},

where \tau_{ij} = \overline{u_i u_j} - \bar{u}_i \bar{u}_j is the subgrid-scale (SGS) stress tensor, \bar{p} is the filtered pressure, \nu is the kinematic viscosity, and the incompressibility condition \frac{\partial \bar{u}_j}{\partial x_j} = 0 has been used to simplify the advective term. The viscous term on the right-hand side can be rewritten in conservative form as \frac{\partial}{\partial x_j} \left( \nu \bar{u}_i \frac{\partial \bar{u}_i}{\partial x_j} \right) - \nu \frac{\partial \bar{u}_i}{\partial x_j} \frac{\partial \bar{u}_i}{\partial x_j}, where the second part represents the resolved viscous dissipation -\nu \bar{S}_{ij} \bar{S}_{ij} and \bar{S}_{ij} = \frac{1}{2} \left( \frac{\partial \bar{u}_i}{\partial x_j} + \frac{\partial \bar{u}_j}{\partial x_i} \right) is the resolved strain-rate tensor. Similarly, the pressure term becomes -\frac{\partial (\bar{p} \bar{u}_j)}{\partial x_j}, and the SGS term expands to -\frac{\partial (\bar{u}_i \tau_{ij})}{\partial x_j} + \tau_{ij} \frac{\partial \bar{u}_i}{\partial x_j}. Thus, the full resolved kinetic energy equation consists of temporal and spatial transport terms (advection, pressure work, viscous diffusion, and SGS flux divergence) balanced by the resolved viscous dissipation and the interscale energy transfer term \Pi_r = \tau_{ij} \bar{S}_{ij}. The key term \Pi_r governs the production of resolved kinetic energy K_r through interactions with the SGS scales, often denoted as the SGS dissipation \varepsilon_{sgs} = -\tau_{ij} \bar{S}_{ij}. In typical turbulent flows, \varepsilon_{sgs} > 0 represents a forward cascade where energy drains from resolved large eddies to unresolved small scales, consistent with Kolmogorov's theory of turbulence. However, \Pi_r can locally become positive, indicating backscatter—a reverse transfer of energy from SGS to resolved scales—which introduces variability and can stabilize or destabilize the flow depending on the context. This term underscores LES's role in explicitly resolving large-eddy production (via mean-flow interactions and resolved-scale transfers) while modeling the small-scale dissipative drain. For compressible flows, the filtered kinetic energy equation extends to include density variations and additional dissipation mechanisms. The resolved kinetic energy becomes \frac{1}{2} \overline{\rho} \tilde{u}_i \tilde{u}_i / \bar{\rho}, and the equation incorporates terms from the filtered continuity and energy equations, such as dilatational dissipation -\nu (\frac{\partial \tilde{u}_k}{\partial x_k})^2 arising from compressibility effects, along with contributions from filtered internal energy transport and SGS heat flux. These extensions account for acoustic wave propagation and shock interactions prevalent in high-speed flows. The implications of this equation are central to LES validation through budget analyses, which quantify how well the simulation conserves energy across scales. By comparing resolved production, transfer, and dissipation terms against direct numerical simulation (DNS) data or experimental measurements, researchers assess model accuracy in preserving the turbulent energy cascade without artificial accumulation or depletion. Such analyses reveal deficiencies in SGS closures if backscatter is underrepresented or if compressible terms lead to unphysical heating.^[12]

Subgrid-Scale Modeling

Functional Models

Functional models in large eddy simulation approximate the subgrid-scale (SGS) stress tensor by drawing an analogy to the effects of molecular viscosity in the Navier-Stokes equations, treating the unresolved scales as an effective eddy viscosity that dissipates energy at small scales.^[13] These models are phenomenological in nature, relying on dimensional analysis and assumptions about the local equilibrium between the production and dissipation of turbulent kinetic energy at subgrid scales, rather than deriving closures from first principles.^[13] The foundational functional model is the Smagorinsky eddy-viscosity model, introduced in 1963 for simulating atmospheric flows. It models the anisotropic part of the SGS stress tensor as

\tau_{ij} - \frac{1}{3} \tau_{kk} \delta_{ij} = -2 (C_s \Delta)^2 |\tilde{S}| \tilde{S}_{ij},

where \tau_{ij} is the SGS stress, \Delta is the filter width, \tilde{S}_{ij} is the resolved strain-rate tensor, |\tilde{S}| = \sqrt{2 \tilde{S}_{ij} \tilde{S}_{ij}} is its magnitude, and C_s \approx 0.18 is an empirically determined constant. This formulation assumes that the SGS stresses align with the resolved strain, leading to an isotropic eddy viscosity \nu_t = (C_s \Delta)^2 |\tilde{S}|.^[13] To address the fixed coefficient's limitations across flow regimes, the dynamic Smagorinsky model was developed, which computes a scale- and flow-dependent C_s using information from both the grid filter and a coarser test filter. The Germano identity relates the resolved stresses at the test-filter level to the SGS stresses at the grid level, expressed as T_{ij} - \hat{\tau}_{ij} = L_{ij}, where T_{ij} and \hat{\tau}_{ij} are test-filtered SGS stresses, and L_{ij} is the resolved Leonard stress.^[14] The model coefficient is then optimized via least-squares minimization of the error in this identity, with Lilly's 1992 normalization ensuring stability by averaging the coefficient over homogeneous directions and clipping negative values.^[15] The wall-adapting local eddy-viscosity (WALE) model improves near-wall behavior by basing the eddy viscosity on the square of the velocity gradient tensor, g_{ij} = S_{ij}^d : S_{ij}^d - \frac{1}{12} (S_{ik}^d S_{jk}^d + S_{jk}^d S_{ik}^d) \delta_{ij}, where S_{ij}^d is the traceless symmetric part of the gradient.^[16] This yields \nu_t = (C_w \Delta)^2 \frac{(g_{ij} g_{ij})^{3/2}}{(S_{ij} S_{ij})^{5/2} + (g_{ij} g_{ij})^{5/4}} with C_w \approx 0.325, enabling proper y^3 scaling of \nu_t in the viscous sublayer without requiring ad hoc damping functions.^[16] Despite their widespread use, functional models like Smagorinsky exhibit limitations, including overdissipation in transitional flows where they fail to distinguish laminar from turbulent regions, leading to premature turbulence suppression.^[17] They also assume local equilibrium between SGS production and dissipation, which does not hold in non-equilibrium flows such as those with strong shear or buoyancy.^[18] Evaluation of these models occurs through a priori tests, which compare modeled SGS stresses against those extracted from direct numerical simulation (DNS) data without solving the full LES equations, and a posteriori tests, which assess the complete LES predictions against reference data for integrated quantities like mean velocities and turbulence statistics.^[18] For instance, a priori analyses reveal that dynamic variants better capture backscatter in DNS fields, while a posteriori simulations demonstrate improved agreement with experiments in channel flows compared to static models.^[19]

Structural Models

Structural models in large eddy simulation (LES) approximate subgrid-scale (SGS) stresses by exploiting the similarity between resolved and subgrid scales, rather than relying on phenomenological analogies like eddy viscosity. These models aim to reconstruct or mimic the interactions between scales using operations on the filtered velocity field, such as additional filtering or deconvolution, to capture the structural features of turbulence more accurately. Unlike functional models, structural approaches can inherently account for both forward and reverse energy transfers, including nonlocal effects and backscatter from small to large scales. The scale-similarity hypothesis forms the foundation of many structural models, proposing that the SGS stress tensor \tau_{ij} can be approximated by the Leonard stress at the test-filter level,

\tau_{ij} \approx \hat{(\bar{u}_i \bar{u}_j)} - \hat{\bar{u}}_i \hat{\bar{u}}_j,

where \bar{\cdot} denotes the grid filter and \hat{\cdot} a coarser test filter applied to the resolved field and its product. This approximation assumes that the smallest resolved scales exhibit dynamics similar to the subgrid scales, enabling a direct estimation of SGS contributions without additional parameters. Introduced by Bardina et al. in 1980, this model demonstrated improved correlation with exact SGS stresses in homogeneous isotropic turbulence compared to earlier approaches, though it tends to underpredict the magnitude of stresses. To enhance the predictive capability of scale-similarity models, mixed models combine them with eddy-viscosity terms, such as the Smagorinsky model, often in a clipped or dynamic formulation to balance scale similarity with dissipative effects. In the dynamic mixed model proposed by Zang et al. in 1993, the scale-similarity component explicitly computes the Leonard stress (arising from filtering the nonlinear term) and cross-stress, while the eddy-viscosity models the remaining SGS Reynolds stress, leading to better agreement with direct numerical simulation data in turbulent recirculating flows like lid-driven cavities. This hybrid approach improves accuracy in regions of strong strain and reduces excessive dissipation, with the dynamic procedure adjusting coefficients locally based on scale-invariance assumptions. Deconvolution methods advance structural modeling by approximating the inverse of the filter operator to recover an estimate of the unfiltered velocity field from the resolved one, thereby computing SGS stresses more precisely. Techniques include Taylor series expansions around the filtered field or iterative approximate deconvolution models (ADM), which use a series of successive filterings to iteratively refine the reconstruction. Developed by Stolz and Adams in the late 1990s and early 2000s, the ADM applies a deconvolution operator G such that the approximate unfiltered velocity is u_i^* = G \bar{u}_i, and the SGS stress is then \tau_{ij} \approx u_i^* u_j^* - \bar{u}_i \bar{u}_j, showing superior performance in wall-bounded flows by capturing near-wall physics without explicit viscosity. These methods, reviewed extensively by Sagaut, provide a framework for handling complex filter kernels and have been extended to compressible flows. Post-2015 developments have incorporated data-driven techniques, particularly neural networks, to learn structural SGS models from high-fidelity simulations like direct numerical simulations (DNS). These approaches train deep neural networks on pairs of filtered and unfiltered data to predict SGS stresses or scale interactions, emphasizing nonlocal and anisotropic effects that traditional models overlook. For example, studies using convolutional neural networks have shown improved predictions of SGS effects in turbulent flows, including backscatter. Recent advances as of 2025 include recursive neural-network-based models for homogeneous isotropic turbulence and multiscale convolutional networks that enhance generalization across Reynolds numbers and flow regimes, offering a pathway to physics-informed structural modeling without hand-crafted assumptions.^[9]^[20]^[21] Overall, structural models excel in capturing backscatter—energy transfer from subgrid to resolved scales—and nonlocal interactions, which functional models often suppress through forward-scatter assumptions, leading to more realistic turbulence statistics in transitional and homogeneous flows. This capability enhances the fidelity of LES in applications requiring precise scale coupling, though computational overhead from additional filtering or training remains a challenge.

Numerical Implementation

Discretization and Filter Application

In large eddy simulation (LES), spatial discretization of the filtered Navier-Stokes equations is typically achieved using finite difference, finite volume, or spectral methods, selected based on the flow geometry and required accuracy. Finite difference methods, often employing high-order schemes such as 4th-order compact differencing, are favored for their low dispersion errors, which are critical for resolving turbulent structures without introducing artificial wave propagation in high-frequency modes. These schemes minimize numerical dissipation while maintaining stability, as demonstrated in simulations of isotropic turbulence where scales with fewer than four points per wavelength are effectively damped. Spectral methods, including spectral element approaches, offer exponential convergence for smooth flows and are particularly effective for periodic domains or global simulations, enabling efficient resolution of large-scale eddies through modal decompositions. Finite volume methods, with their conservation properties, are commonly used for complex geometries, incorporating flux limiters to handle shocks or discontinuities in compressible flows. Temporal integration in LES balances accuracy with stability, often employing explicit schemes like multi-stage Runge-Kutta methods for their simplicity and low cost per step, suitable when subgrid-scale (SGS) terms are not overly stiff. For instance, a 5-stage explicit Runge-Kutta scheme has been applied to jet flow simulations, ensuring second-order accuracy while adhering to the Courant-Friedrichs-Lewy (CFL) condition, which constrains the time step Δt ≤ Δx / |u| to prevent numerical instability from convective terms. Implicit methods, such as backward Euler or Crank-Nicolson, are preferred for stiff problems involving dominant viscous or SGS effects, allowing larger time steps but increasing computational expense due to iterative solvers. The CFL condition remains a key limiter, typically requiring CFL numbers below 0.5-1.0 for explicit schemes in turbulent flows to capture unsteady dynamics accurately. Filter application in discrete LES can be explicit or implicit, directly influencing the representation of unresolved scales. Explicit filtering involves convolving the filtered variables with a kernel (e.g., Gaussian or top-hat) on the computational grid, often using high-order compact filters with a 5-point stencil to approximate the continuous filter operator while preserving symmetry and accuracy; the filter width Δ is dynamically adjusted to local grid resolution, such as Δ = (Δx Δy Δz)^{1/3} for nonuniform meshes. Implicit filtering, conversely, arises from the numerical scheme's inherent dissipation, where the truncation error acts as a low-pass filter without additional computational overhead. In implicit LES (ILES), high-resolution shock-capturing schemes like weighted essentially non-oscillatory (WENO) or total variation diminishing (TVD) methods, combined with monotonicity-preserving upwind biasing, implicitly model SGS stresses through numerical viscosity, eliminating the need for explicit closures and enabling robust simulations of transitional or multiphase flows. Grid requirements for LES emphasize resolving the energy-containing large eddies while modeling smaller scales, generally necessitating 10-20 grid points per largest eddy wavelength to capture at least 80-90% of the turbulent kinetic energy. For wall-bounded flows, this translates to grid spacings of Δx^+ ≈ 50-100 in streamwise and spanwise directions and Δy^+ ≈ 1 near walls for well-resolved LES, with total grid points scaling as Re^{13/7} to Re^{9/4} depending on the Reynolds number and modeling approach. Adaptive mesh refinement (AMR) enhances efficiency for multiscale flows by locally increasing resolution in regions of high vorticity or strain, using indicators like Q-criterion or velocity gradients to refine cells, thereby reducing overall computational cost by 50-80% in applications such as combustor simulations while maintaining filter consistency across levels.

Boundary Conditions

In large eddy simulation (LES), boundary conditions play a critical role in capturing realistic turbulent flows, particularly near interfaces where artificial reflections or inconsistencies can distort the resolved scales. Proper treatment ensures that the simulation domain interacts appropriately with external conditions, maintaining turbulence statistics and flow physics without introducing numerical artifacts. For inflow boundaries, realistic turbulent fluctuations must be prescribed to initiate and sustain turbulence within the domain. Synthetic turbulence generation methods, such as the digital filter approach, create spatially and temporally correlated velocity fields by applying filters to white noise, allowing specification of turbulence intensities and length scales for practical implementation.^[22] Alternatively, precursor simulation recycling draws from a separate equilibrium turbulent simulation to map data onto the inlet, ensuring compatibility with the Navier-Stokes equations and natural turbulence spectra, as introduced by Lund et al. (1998).^[23] These methods address the challenge of generating grid-scale turbulence without excessive computational overhead. At wall boundaries, the near-wall region demands careful resolution due to the dominance of small-scale structures. In wall-resolved LES, the grid is refined to y⁺ < 1, directly resolving the viscous sublayer and requiring streamwise and spanwise resolutions of approximately Δx⁺ ≈ 50 and Δz⁺ ≈ 15 to capture near-wall streaks and vortices. This approach, however, scales poorly with Reynolds number, demanding O(Re_L^{9/4}) grid points. In contrast, wall-modeled LES employs empirical models, such as the logarithmic law or thin-boundary-layer equations, to approximate subgrid-scale effects in the inner layer, enabling coarser grids with O(Re_L) points and y* ≈ 0.1–0.2δ matching between resolved and modeled regions. Near walls, subgrid-scale modeling must account for anisotropic stresses, often integrating wall functions briefly referenced from structural models. Outflow boundaries require conditions that allow disturbances to exit the domain without upstream propagation. Convective boundary conditions, such as the Orlanski radiation scheme, extrapolate waves at their phase speed to minimize reflections, with a hybrid formulation incorporating a minimum speed tied to the normal velocity.^[24] Non-reflecting approaches, including Sommerfeld-type conditions, apply homogeneous Neumann constraints to scalars like temperature while adjusting for buoyancy in stable flows, ensuring minimal influence on interior turbulence.^[25] In canonical setups like channel flows, no-slip conditions enforce zero velocity at solid walls to model viscous effects accurately, contrasting with periodic boundaries in streamwise and spanwise directions that assume homogeneity and repeat flow patterns indefinitely.^[26] This combination sustains fully developed turbulence without inlet/outlet artifacts, though it limits applicability to non-homogeneous geometries. Key challenges in LES boundary treatments include maintaining mass conservation across interfaces, often via correction factors in open boundaries, and ensuring long-term turbulence sustainability, as poor inflow coherence can lead to artificial decay or instability.^[25] These issues are exacerbated in complex domains, necessitating validation against experimental data to avoid spurious reflections or imbalances.

Applications and Challenges

Practical Applications

Large eddy simulation (LES) has been extensively applied in aerodynamics to predict complex turbulent flows around aircraft components. In simulations of wing-body junctions, such as the NASA Juncture Flow experiment, LES captures the separation and vortex structures at the wing-fuselage juncture, validating against experimental data for trailing-edge separation progression under subsonic conditions.^[27] For bluff body flows, LES resolves the vortex shedding behind circular cylinders, accurately predicting Strouhal numbers and wake dynamics in the near field, which is critical for understanding drag and noise in engineering designs.^[28] In combustion engineering, LES models premixed and non-premixed flames in gas turbine combustors by resolving large-scale turbulent structures while subgrid models handle flame wrinkling and propagation. For instance, in General Electric's LM6000 lean-premixed combustor, LES simulates swirl-stabilized flames, capturing heat release fluctuations and thermoacoustic instabilities associated with flame dynamics.^[29] Similarly, Siemens Energy employs LES to investigate transient combustion effects in their engines, revealing mechanisms of thermoacoustic vibrations through detailed resolution of flame interactions with turbulence.^[30] Environmental flows benefit from LES in simulating atmospheric boundary layers (ABL) for wind energy applications, where actuator line models represent turbine blades to predict wake interactions and power output in farm configurations. These simulations incorporate ABL turbulence to assess turbine array performance under varying stability conditions.^[31] In urban settings, LES evaluates pollutant dispersion from traffic sources, resolving street canyon flows and scalar transport to map concentration hotspots and inform air quality management.^[32] Multiphase flows, such as particle-laden jets, are analyzed using LES coupled with Euler-Lagrange methods to track particle trajectories and turbulence modulation in dilute suspensions, aiding design of industrial sprays.^[33] For bubbly flows in chemical reactors, LES predicts bubble-induced turbulence and mixing in bubble columns, capturing coalescence and breakup effects for reactor optimization.^[34] Recent advances in the 2020s extend LES to hypersonic re-entry vehicles, where it simulates turbulent boundary layers and shock interactions around capsules like Schiaparelli, quantifying aerothermal loads during parachute deployment.^[35] In bio-fluids, LES models transitional blood flow in arteries, such as the thoracic aorta, resolving pulsatile turbulence in the healthy thoracic aorta to evaluate characteristics like wall shear stress variations and their potential links to vascular diseases.^[36]

Advantages and Limitations

Large eddy simulation (LES) offers significant advantages over Reynolds-averaged Navier-Stokes (RANS) methods by directly resolving the unsteady dynamics and anisotropy of large-scale turbulent eddies, which are critical for accurately predicting complex flows such as those involving separation or transition.^[1] Unlike RANS, which relies on time-averaging and empirical closures that often fail in non-equilibrium conditions, LES captures these dominant structures explicitly, leading to improved fidelity in simulations of three-dimensional and separated flows.^[1] Additionally, LES bridges the gap between direct numerical simulation (DNS), which resolves all scales but is prohibitively expensive, and RANS, providing a favorable balance of accuracy and computational feasibility for practical engineering problems.^[37] The method's scalability to complex geometries is enhanced through hybrid approaches, such as detached eddy simulation (DES), which combine LES in the outer flow with RANS near walls to handle intricate boundaries efficiently without excessive grid refinement.^[38] However, LES incurs a substantial computational cost, scaling as O(\mathrm{Re}^{9/4}) with the Reynolds number \mathrm{Re}, compared to DNS's steeper O(\mathrm{Re}^3) dependence, necessitating massive parallel computing resources—often petascale systems—for industrial-scale applications like turbine or combustor flows.^[37]^[39] Key limitations include high sensitivity to subgrid-scale (SGS) models and filter choices, which can introduce errors in non-homogeneous or transitional flows if not calibrated properly, as static models like Smagorinsky's underperform in such cases while dynamic variants offer better adaptability but at added complexity.^[1] Grid resolution requirements are demanding, typically spanning $10^6 to $10^9 cells for wall-modeled LES in high-Re industrial cases, with inadequate resolution leading to excessive numerical dissipation or unresolved small-scale interactions.^[40] Furthermore, challenges persist in low-Re regimes and laminar-turbulent transitions, where SGS modeling struggles to capture intermittent physics accurately.^[1] In comparison to RANS, which provides steady, statistically averaged solutions at lower cost but misses unsteady features, LES excels in transient predictions yet demands orders-of-magnitude more resources; against DNS, it sacrifices small-scale resolution for feasibility but enables simulations at practically relevant Re up to $10^6 or higher.^[37] Hybrid LES/RANS strategies address wall-layer inefficiencies by applying RANS in near-wall regions, reducing overall grid points while preserving LES accuracy in the core flow.^[38] Looking ahead, future challenges for LES include advancing uncertainty quantification to assess modeling and numerical errors in high-dimensional simulations, integrating data assimilation techniques to incorporate experimental observations and reduce predictive variability, and optimizing efficiency on exascale hardware to handle multiphysics and real-time industrial workflows post-2020.^[41]^[42]^[39]

References

[1]
[PDF] Large-eddy simulation: achievements and challenges
In this paper, the present state of the large-eddy simulation (LES) technique is discussed. Modeling and numerical issues that are under study will be ...<|control11|><|separator|>
[2]
[PDF] Mathematical Perspectives on Large Eddy Simulation models for ...
The main objective of this paper is to review and report on key mathematical issues related to the theory of Large Eddy Simulation of turbulent flows. The ...
[3]
GENERAL CIRCULATION EXPERIMENTS WITH THE PRIMITIVE ...
Abstract. An extended period numerical integration of a baroclinic primitive equation model has been made for the simulation and the study of the dynamics ...
[4]
Large Eddy Simulation for Incompressible Flows - SpringerLink
Mar 9, 2013 · Large Eddy Simulation, where the small-scale turbulent fluctuations are themselves smoothed out and modelled via eddy-viscosity and diffusivity assumptions.
[5]
Large-eddy simulation: Past, present and the future - ScienceDirect
An alternative approach is called large-eddy simulation (LES) which was proposed in as early as 1963 by Smagorinsky. ... This paper presents briefly LES formalism ...
[6]
More for LES: A Brief Historical Perspective of MILES (Chapter 1)
The subject of this chapter is monotone integrated large eddy simulation (LES), or MILES – monotonicity-preserving implicit LES (ILES), a class of practical ...
[7]
Review Status and future developments of Large-Eddy Simulation of ...
The contribution presents the possibilities offered today by turbulent scale-resolving strategies (Large-Eddy Simulation, LES) to treat complex, multiphase flow ...Missing: growth | Show results with:growth
[8]
Deep Learning for Subgrid‐Scale Turbulence Modeling in Large ...
May 10, 2022 · Deep learning-based subgrid-scale models outperform traditional eddy-viscosity models ... Deep neural networks for data-driven LES closure models.Abstract · Introduction · Methodology · Results
[9]
The Energy Exascale Earth System Model Version 3: 1. Overview of ...
Oct 10, 2025 · This paper describes the atmospheric component of the US Department of Energy's Energy Exascale Earth System Model (E3SM) version 3.
[10]
[PDF] energy cascade in large-eddy simulations of turbulent fluid flows
In this model, the energy cascade is then viewed solely as an energy loss of the large-scales due to an artificial viscosity arising from subgrid-scale motions.
[11]
[PDF] Explicitly filtered large eddy simulation on unstructured grids
Nominally, the local filter width is estimated to be 1.8 times the local grid size. A more rigorous estimate would require the solution of the Green's ...
[12]
Kinetic-energy-flux-constrained model using an artificial neural ...
Dec 3, 2021 · In the filtered kinetic energy equation (3.1), the KEF \varPi _{\varDelta } is an unclosed term and needs to be modelled. Based on the tensor ...
[13]
Smagorinsky Model - an overview | ScienceDirect Topics
The Smagorinsky model is defined as a pioneering subgrid-scale model for large eddy simulation (LES) that uses a grid-characteristic length scale and a ...<|control11|><|separator|>
[14]
[PDF] A dynamic subgrid-scale eddy viscosity model
Additional modifications to the Smagorinsky model were made in the near-wall region of plane channels to force the subgrid-scale stresses to vanish at the solid ...
[15]
[PDF] A proposed modification of the Germano subgrid-scale closure method
Smagorinsky,” hence S, closure. The S closure was first applied extensively to three-dimensional turbulence simu- lations by Deardorff.3 Lilly4 had earlier ...Missing: normalization | Show results with:normalization
[16]
[PDF] Subgrid-scale stress modelling based on the square of the velocity ...
Nov 28, 2013 · The resulting WALE (Wall-Adapting Local Eddy-viscosity) model is based on a tensor invariant and reproduces the proper scaling at the wall (νt = ...Missing: original | Show results with:original
[17]
[PDF] Predicting natural transition using large eddy simulation
The constant coefficient Smagorinsky (CCS) model fails to differentiate between lam- inar and turbulent flow causing the turbulent eddy viscosity to be active ...
[18]
A priori and a posteriori evaluations of sub-grid scale models for the ...
Nov 5, 2016 · This paper focuses on five of them: the static Smagorinsky model (SS) [3], [4], the dynamic Smagorinsky model (DS) [5], the scale-similarity ...
[19]
[PDF] A priori and a posteriori evaluations of sub-grid scale models for the ...
In the present study, we perform a priori and a posteriori evaluations of five SGS ... In this case, different SGS models produce nearly the same LES solution as ...
[20]
https://www.cambridge.org/core/journals/journal-of-fluid-mechanics/article/recursive-neuralnetworkbased-subgridscale-model-for-large-eddy-simulation-application-to-homogeneous-isotropic-turbulence/1271EA1EF4678BBDA6A2D9B14BFDE1AE
[21]
https://arxiv.org/abs/2502.10814
[22]
https://doi.org/10.1016/j.compfluid.2009.10.007
[23]
https://doi.org/10.1006/jcph.1998.5882
[24]
Open boundary conditions for atmospheric large-eddy simulations ...
May 17, 2024 · Open boundary conditions were developed for atmospheric large-eddy simulation (LES) models and implemented into the Dutch Atmospheric Large-Eddy Simulation ...
[25]
[PDF] LARGE EDDY SIMULATION IN A CHANNEL No KO Mitra
The difference between the fully developed turbulent channel flow obtained with LES with periodicity condition and the inlet and exit and the LES with fully ...
[26]
[PDF] Wall-modeled LES of the NASA Juncture Flow Experiment
We explored the performance of Wall-modeled Large-Eddy Simulation (WMLES) using an unstructured finite volume solver and equilibrium wall model for the NASA ...
[27]
A Large-Eddy Simulation of the Near Wake of a Circular Cylinder
The formation and the downstream transport of the Strouhal vortices in the near wake of a circular cylinder are investigated using the large-eddy simulation ...
[28]
LES studies of the flow in a swirl gas combustor - ScienceDirect.com
Here, we examine the flow and the flame in a model gas turbine combustor (General Electric's lean premixed dry low NOx LM6000) to evaluate the potential of LES ...
[29]
Siemens Energy uses LES simulation to investigate transient ...
Large eddy CFD simulations reveal causes of thermoacoustic vibrations. Siemens Energy uses LES simulation to investigate transient combustion effects ...
[30]
Large-eddy simulation of atmospheric boundary layer flow through ...
This paper focuses on recent research efforts to develop and validate an LES framework for wind energy applications.
[31]
Large-Eddy Simulation of Airflow and Pollutant Dispersion in ... - MDPI
Jun 28, 2022 · The atmospheric flow and dispersion of traffic exhaust were numerically studied in this work while considering a model street canyon intersection of a city.
[32]
Applicability of large eddy simulations to capture turbulence ...
Feb 10, 2022 · The main focus of those investigations was to perform LES to predict the particle dynamics and concentration profiles in particle-laden flows.
[33]
Eulerian–Lagrangian based large-eddy simulation of a partially ...
Gas–liquid bubbly flow in a flat bubble column (“Becker” case with a gas flow rate of 1.6 l/min) is studied by means of large-eddy simulation (LES) combined ...
[34]
Large-Eddy Simulations of a Hypersonic Re-Entry Capsule Coupled ...
The goal of this paper is to investigate the aerodynamic and aerothermodynamic behavior of the Schiaparelli capsule after the deployment of a supersonic disk- ...
[35]
Characteristics of transition to turbulence in a healthy thoracic aorta ...
Jan 25, 2025 · This study employed large eddy simulation (LES) with the wall-adapting local eddy-viscosity (WALE) model to investigate transitional flow ...Missing: arteries 2020s
[36]
Wall-Modeled Large-Eddy Simulation for Complex Turbulent Flows
We review recent investigations in wall-modeled LES, including the development of novel approximate boundary conditions and the application of wall models to ...
[37]
Advances and challenges of applied large-eddy simulation
Aug 10, 2025 · Large eddy simulation (LES) can predict unsteady flows accurately but necessitates high-resolution meshes in the boundary layer and substantial ...
[38]
[PDF] Petascale Computing in Europe
Scientific Case for European Petascale Computing. 79. Engineering. Large eddy simulation of the ignition of a helicopter gas turbine. Two hot gas streams (red) ...
[39]
(PDF) Grid-point requirements for large eddy simulation: Chapman's ...
Aug 7, 2025 · A review is presented of estimates for the grid resolution requirements for wall-resolved and wall-modeled large-eddy simulation, of flows with ...
[40]
Assessing uncertainties from physical parameters and modelling ...
Mar 29, 2021 · In this study, we investigate uncertainties in a large eddy simulation of the atmosphere, employing modern uncertainty quantification methods.Missing: challenges assimilation exascale
[41]
Ensemble-variational assimilation of statistical data in large-eddy ...
Oct 21, 2021 · A nonintrusive data assimilation methodology is developed to improve the statistical predictions of large-eddy simulations (LES).Missing: exascale | Show results with:exascale<|control11|><|separator|>