CFD-101: The Basics of Computational Fluid Dynamics Modeling
Reynolds Number Restrictions in CFD
What are the highest and lowest Reynolds number flows that can be accurately computed by a given numerical method? This question has a variety of answers, and, as with most technical issues, the variety of answers arises from the assumptions involved in giving the answer.
For present purposes, the Reynolds number R is defined as R=LU/ν, where L and U are characteristic length and velocity scales for a flow, and ν is the kinematic viscosity of the fluid. It will be recalled that the non dimensional Reynolds number is a measure of the importance of inertia to viscosity effects. At high Reynolds numbers a flow may become turbulent, exhibiting qualitatively different behavior.
Generally, the most important limit to consider is that of high Reynolds numbers. This is the limit where computations might be used to predict the breakdown of a laminar flow into turbulence, or the lift and drag of a body that is dependent on where boundary layers separate from its surface. In these or other types of flow processes in which it is critical to correctly simulate the relative effect of viscous stresses on the flow, it is useful to have some idea of what level of accuracy can be expected in a computation.
The reason that a Reynolds number limitation exists in computational fluid dynamics CFD) is that the computational stability of most CFD methods relies on some type of numerical smoothing or homogenizing within the computational elements. Since viscosity is a physical mechanism for smoothing flow variations, there can be a problem differentiating between numerical and physical smoothing. This is especially important when critical Reynolds number situations are encountered, because they require an especially accurate estimate of viscous stresses.
High Reynolds Number Limit – A Physical Argument
A simple physical argument can be used to estimate the computational requirements (i.e., resolution) needed to achieve an accurate representation of a flow. The argument is based on the assumption that when a flow region is subdivided into small elements all flow quantities within an element are slowly varying. This assumption carries the implication that the average values of quantities in each element are good approximations for the actual values within the element.
To have a slowly varying velocity within an element, the Reynolds number of the flow on scales of the element size must be small, say of order one, Rd=dx·du/ν ≤ 1.0. In this expression dx and du are length and velocity scales characteristic of the element. This physical requirement, the smoothness of the flow in elements (i.e., a low Reynolds number, laminar flow on this scale), may be used to define the size of elements needed for an accurate numerical resolution.
The above inequality can be converted to a macroscopic Reynolds number by the relations, L=Ndx and U=Ndu, which leads to R ≤ N2. In other words, the physical accuracy requirement of a smooth flow on the scale of individual elements implies that the maximum Reynolds number one can expect to compute with accuracy is on the order of NN2 where N is the number of elements used to resolve a characteristic length L.
In typical applications, N is often in the range of 10 to 20, which translates to a maximum Reynolds number for accurate computations of only about 400, not a very large number! Before commenting on this result it is instructive to try a different approach for estimating the limit for accurate Reynolds number computations.
High Reynolds Number Limit – A Numerical Argument
The amount of viscous-like smoothing introduced into a computation by numerical approximations can be estimated from truncation errors. The idea is to do a Taylor Series expansion on the difference approximations in powers of the element size (and time-step size if that is appropriate). Of course, a consistent approximation should have as its lowest order terms the partial differential equation that was originally being approximated.
At the next higher order there are usually terms that have the character of a diffusion (i.e., second-order space derivatives). A comparison of the coefficients of these terms with the coefficient of viscosity gives an estimate of when viscous effects would no longer be computed accurately.
For a first-order numerical approximation (e.g., a donor cell or upwind technique for advection) the ratio of terms, which must be less than one for accuracy, leads to the criteria R ≤ 2N. With a second-order approximation the result is R ≤ N2, the same result obtained from the "Physical Argument."
There are small numerical factors multiplying the right-hand sides of these relations, which depend on the specific numerical approximations used, but the basic dependencies on N remain unchanged. Any second-order method is clearly much better than a first-order method, but the results are not encouraging. The maximum Reynolds number that can be computed accurately appears to be quite limited, unless one is willing to increase N, which means dealing with extremely large grids.
General Comments on High Reynolds Numbers
These estimates are discouraging when first encountered, but there are frequently mitigating circumstances. Foremost is the realization that most problems do not require an accurate treatment of viscous stresses. For these problems the high Reynolds number limit has the intended meaning that viscous effects are not important.
When flows have a high enough Reynolds number to be fully turbulent the momentum mixing induced by the turbulence often leads to a mean flow with an effective Reynolds number that is less than 100, well within the range of resolvable scales. Of course, this assumes that a suitable turbulence model is available to describe the turbulence.
Finally, when it is necessary to have some flow property that depends on an accurate knowledge of viscous effects, it may be possible to induce that effect by artificial means. For example, in wind tunnels trip wires are sometimes used to trigger flow separations to account for a lack of Reynolds number similarity. A similar treatment can be added to a numerical simulation of a wind tunnel.
The bottom line is, CFD methods can be used to compute high Reynolds number flows, but it is up to the modeler to be alert for situations where numerical errors could overshadow physical effects.
Low Reynolds Number Limit
At low Reynolds numbers the limit is not one of accuracy but a limit based on the computational time necessary to complete a computation. When explicit numerical approximations are used for viscous stress terms there is a limit on the size of the time step to maintain numerical stability. That limit is essentially a statement that momentum changes caused by viscosity do not propagate more than about one element in one time step. In a simple two-dimensional case this limit is νdt ≤ dx2/4.
This can be transformed into an expression involving the Reynolds number by making the correspondences: T=Mdt and TU=L. That is, the characteristic time for a flow is the time for fluid at velocity U to move a distance L, and the number of time steps resolving time T is M. With these relations the stability condition is then, M = 4N2/R.
The importance of this result is that M increases inversely with R. For very low Reynolds number flows, explicit numerical methods may require a very large number of time steps, and this number increases rapidly with an increase in resolution. The low Reynolds number limit is best eliminated by employing an implicit numerical method for evaluating viscous stresses.