During the past decades turbulence models of various complexities have been developed. Turbulence models that employ the most assumptions are typically the least demanding from a CPU cost standpoint.

As the number of assumptions associated with the model decreases, the cost and accuracy of the model typically increases. Since no universal turbulence model exists, CFD users need to choose the best turbulence model specific to their application. The most cost effective approach is typically steady Reynolds-averaged Navier-Stokes (RANS). When the flow is naturally unsteady and exhibits oscillations in the quantities of interest, URANS may be employed to improve the accuracy of solutions and gain a better understanding of how much the flow field fluctuates as a function time. For applications when the flow transitions from laminar to turbulent as a function of space, a dedicated transition model is required in order to accurately predict the near wall behavior. When the application demands additional knowledge of the turbulent fluctuations for separated, reattaching or impinging flow fields, hybrid simulations (Detached Eddy Simulation (DES), Delayed Detached Eddy Simulation (DDES)) may be employed. Although the modeling assumption is still made within the boundary layer, the tradeoff between accuracy and computational efficiency is often worthwhile compared to Large Eddy Simulation (LES) or Direct Numerical Simulation (DNS). Finally, when the application requires detailed information about nearfield fluctuations, or for combustion and acoustic applications, LES should be considered.

Based on the information presented it is up to you to understand the complexity of your simulation and make an appropriate decision on which turbulence model to use. Generally, it is recommended to start with a simple turbulence model since a majority of industrial problems can be well defined with a steady state assumption. Even if the application is inherently unsteady, this is still a good place to start as a significant amount of information can be garnered from running RANS. If the initial RANS fails to provide the level of accuracy required for the specific application you will need to consider the use of a hybrid model or LES. Often times the flow regime exhibits sensitivities to the selection of the turbulence model, specifically RANS vs. Hybrid/LES. In any case, the selection of the turbulence model should be verified against previous simulation or validated with experimental/analytical data for the flow in question.