To simulate fluid flow around a vehicle, a wind turbine, or inside a hydraulic pump, engineers must solve the Navier-Stokes equations. In an ideal world, we would resolve every single eddy—from massive atmospheric swirls down to millimeter-scale dissipative vortices—using Direct Numerical Simulation (DNS). However, because the computational cost of DNS scales cubically with the Reynolds number, simulating real-world engineering systems directly is computationally impossible.
Instead, the engineering community relies on Computational Fluid Dynamics (CFD) powered by turbulence closure models. These models use mathematical approximations to simulate the effects of sub-grid turbulence without resolving every individual vortex.
But there is a catch: no single turbulence model is universally accurate. Every model makes specific assumptions about fluid behavior, and choosing the wrong model can lead to catastrophic errors in predicting lift, drag, heat transfer, or structural stability. Understanding why these models behave differently is the key to selecting the right tool for a given simulation.
The Physics and Intuition Behind Closure Models
The most common engineering approach is the Reynolds-Averaged Navier-Stokes (RANS) framework. RANS splits fluid velocity into a steady mean component and a fluctuating turbulent component. This mathematical split introduces an unsolved term known as the Reynolds stress tensor, which represents how turbulent eddies transport momentum. To solve the flow, we must "close" the equations using a turbulence model.
The models chosen by engineers represent different trade-offs between computational cost and physical fidelity:
The $k$-$\varepsilon$ Model: The Free-Stream Workhorse
The standard or Realizable $k$-$\varepsilon$ model tracks two variables: turbulent kinetic energy ($k$) and the dissipation rate of that energy ($\varepsilon$). It assumes that turbulence is isotropic—meaning it behaves identically in all directions.
- Why it works: It is highly stable and computationally inexpensive, making it excellent for core, high-Reynolds-number flows far away from solid walls.
- Where it fails: It struggles in regions with strong adverse pressure gradients or sharp curvature because its isotropic assumption overpredicts the fluid's resistance to separation.
The $k$-$\omega$ and SST Models: Boundary Layer Specialists
The $k$-$\omega$ model replaces the dissipation rate ($\varepsilon$) with the specific dissipation rate ($\omega$). The Shear Stress Transport (SST) variant is a hybrid: it uses $k$-$\omega$ near solid walls and switches to $k$-$\varepsilon$ in the free stream.
- Why it works: The $k$-$\omega$ formulation handles the low-Reynolds-number physics of the viscous sublayer near walls exceptionally well. By blending this with $k$-$\varepsilon$ in the open domain, the SST model accurately captures boundary layer separation under adverse pressure gradients.
Scale-Resolving and Hybrid Models (LES and DES)
When flows become highly chaotic and transient, time-averaged RANS models break down. Large Eddy Simulation (LES) resolves large, energy-containing eddies directly while modeling only the smallest scales. Detached Eddy Simulation (DES) is a hybrid approach that uses RANS near the walls (where eddies are too small to resolve efficiently) and switches to LES in detached wake regions.
Aerodynamics and the Challenge of Flow Separation
When a fluid encounters an adverse pressure gradient—such as the expanding rear of a vehicle or an aircraft wing at a steep angle—the fluid closest to the surface loses momentum and detaches, creating a turbulent wake. Accurately predicting this separation is the ultimate test for a turbulence model.
k-ε Model (Overpredicts attachment)
===================================> [Delayed Separation / Low Drag Error]
k-ω SST Model (Blends wall/free-stream physics)
===========\
\======================> [Accurate Separation / Correct Lift]
In vehicle design, this choice alters performance metrics. Luwei Wang et al. explored this sensitivity in their study, Analysis of the Influence of Different Turbulence Models on the Prediction of Vehicle Aerodynamic Performance. They noted that while $k$-$\varepsilon$ models are computationally efficient and highly effective for predicting pure aerodynamic drag, $k$-$\omega$ models are required to capture lift accurately, and DES models are essential for capturing highly transient wake dynamics behind the vehicle.
This divergence becomes even more acute at high angles of attack. As shown by Pradeep Kumar Shetty et al. in CFD Simulation of Flow Past MAV Wings, the standard $k$-$\varepsilon$ model fails to capture the massive flow separation that occurs on Micro Air Vehicle (MAV) wings past a critical angle. In contrast, the Spalart-Allmaras (SA) and SST models successfully predict the resulting drop in lift and rise in drag because their underlying equations do not artificially damp out the separation zone.
For complex, bluff-body geometries like a rolling tire, the flow is entirely dominated by detached, unsteadiness. Haichao Zhou et al. evaluated these trade-offs in Evaluation of Shear Stress Transport, Large Eddy Simulation and Detached Eddy Simulation for the Flow around a Statically Loaded Tire. They found that while RANS-based SST models provide a reasonable approximation of wall-bounded forces, a hybrid DES approach is required to capture the intricate, transient vortices shedding into the tire's wake without incurring the prohibitive computational cost of a full LES simulation.
Similarly, in internal aerodynamics and structural safety, model selection dictates design viability:
- Transonic Compressors: Wenhui Yan et al. demonstrated in Numerical Simulation of Transonic Compressors with Different Turbulence Models that the choice of closure model directly changes how shock waves and shock-boundary layer interactions are resolved, altering the predicted efficiency of the compressor.
- Structural Oscillations: When studying galloping instabilities in architectural profiles, M. Chávez et al. found in Numerical Investigation of Galloping Instabilities in Z-Shaped Profiles that comparing SA and SST $k$-$\omega$ models against experimental data was vital to ensure the simulation accurately captured the vortex shedding responsible for dangerous structural oscillations.
Hydrodynamics, Free Surfaces, and Heat Transfer
The sensitivity of turbulence modeling extends far beyond aerodynamics. In hydrodynamics, the interaction between a turbulent fluid and a moving free surface (the boundary between air and water) introduces severe numerical challenges.
When modeling coastal defense structures, Haitao Zhao et al. discovered in Selection of a Turbulence Model for Wave Evolution on a New Ecological Hollow Cube that standard turbulence models often suffer from "unphysical damping," where wave energy artificially dissipates before hitting the structure. They demonstrated that a stabilized $k$-$\omega$ SST model provides the most accurate, grid-independent predictions for wave breaking, run-up, and reflection.
Standard RANS: Wave Energy ===[ Artificial Damping ]===> Weak Impact
Stabilized SST: Wave Energy ============================> Accurate Impact & Breaking
In multiphase engineering, such as ship air lubrication systems designed to reduce hull friction, the turbulence model must resolve how injected air bubbles interact with the boundary layer. Gyeongseo Min et al. showed in Influence of CFD Modelling Parameters on Air Injection Behaviour in Ship Air Lubrication Systems that combining the SST $k$-$\omega$ model with a low-$y^+$ grid configuration (where the mesh is fine enough to resolve the viscous sublayer directly) is essential. This specific combination yields air layer formations and sweep angles that closely match physical sea trials.
For internal flows focused on thermal management, such as cooling channels in gas turbine blades, turbulence is intentionally induced using dimples to enhance heat transfer. In Flow and Heat Transfer in the Tree-Like Branching Microchannel with/without Dimples, Linqi Shui et al. emphasized that a comparative validation of standard models against experimental data is a mandatory first step; choosing a model that miscalculates vortex generation inside the dimples would lead to dangerous underpredictions of local hot spots.
Grid Dependency and the Future of Automated Closures
A turbulence model cannot be separated from the grid on which it is solved. If the mesh is too coarse, even the most sophisticated model will fail. Conversely, some models are highly sensitive to grid alignment and boundary conditions.
This coupling was highlighted by Liying Wang in The Optimal Design based on CFD combined with CAD for Turbine Runner, noting that the accuracy of $k$-$\varepsilon$ optimizations for hydraulic machinery is strictly contingent on grid fidelity. If the grid cannot resolve the geometric curvature, the model's assumptions break down.
The danger of user configuration error was underscored on a massive scale by Shimin Yu et al. in Modelling the hydrodynamic response of a floating offshore wind turbine – a comparative study. In a comparative study involving 17 distinct numerical setups across the field, the authors revealed that even when using identical numerical codes, different choices in meshing, boundary conditions, and numerical schemes produced substantially different hydrodynamic responses. This variability was echoed by Xinkai Li et al. in Analysis of the Effect of Vortex Generator Spacing on Boundary Layer Flow Separation Control, who warned that inappropriate model assumptions can cause simulations of vortex-driven flow control to diverge entirely from real-world behavior.
The Next Horizon: Physics-Informed Machine Learning
To bypass the manual trial-and-error of model selection, researchers are exploring data-driven alternatives. However, purely data-driven neural networks often violate fundamental physical laws, like the conservation of mass or momentum.
Pure Data-Driven NN ===> High Speed ===> Violates Conservation Laws (Unstable)
Physics-Informed NN ===> High Speed ===> Enforces Navier-Stokes Constraints (Stable)
To solve this, William M. Fox et al. investigated a hybrid approach in Optimising Physics-Informed Neural Network Solvers for Turbulence Modelling: A Study on Solver Constraints Against a Data-Driven Approach. They discovered that by embedding carefully reduced fluid dynamic equations directly into the loss function of a Physics-Informed Neural Network (PINN), they could constrain the model to obey physical boundaries. This framework achieves superior accuracy and convergence speed compared to traditional data-driven models, offering a glimpse into a future where machine learning automatically optimizes and refines turbulence closures for complex engineering flows.
Key Foundational Work
CFD Simulation of Flow Past MAV Wings (2013) — contributed 2 concept elements to this direction.
The Optimal Design based on CFD combined with CAD for Turbine Runner (2012) — contributed 1 concept elements to this direction.