With computer-based fluid simulation already being extremely realistic (and likely faster and cheaper than physical wind tunnels), are they, anywhere, completely replacing physical wind tunnels? If not, is this planned anywhere, or even possible?
The thing about computer models is that they are just that; models. They can only approximate the actual behavior we see in the real world.
Computer models are good for trying many different ideas and making predictions. It can be faster to create a new computational test than to make a physical model for a wind tunnel. But very detailed fluid dynamics calculations require large amounts of computing power, and this takes time and money as well.
Wind tunnels can give much more detailed and accurate information, and are useful for validating the results given by the computers. The results don't always match, which means that the computer models must be revised. This is where wind tunnels are still very useful. They can test an entire airplane in actual flow, where computing that detail across the whole aircraft would require long computations.
Although a wind tunnel is also only a model of real conditions, it provides an actual body in actual flow conditions, which means the results can correlate much more closely to the real conditions being tested. Computational analysis is limited by assumptions, the chosen modeling method, and computational power.
Until computing power reaches the point where it can run models that will truly match a wind tunnel's accuracy and detail, there will still be a need to use wind tunnels to provide data.
At the end we are taking about modeling, where we are having computer modeling and on the other side we have hardware modeling. When we are talking about modeling we have 3 levels of modelization:
- Geometry and external conditions modeling: approximation made of the geometry and the external conditions defined in the problem. This includes the supports created in wind tunnel for example.
- Numerical discretization: how the external flow domain is discretized numerically.
- Turbulence modelling: how the external physics is represented (Euler, RANS, LES...)
Hardware modelling is only affected by the first parameter, for example, Reynolds number is not the same in many wind tunnel. Tests and the external conditions are replaced by walls. In fact, in the area of geometry and external conditions, numerical simulation is far more advanced than current WTT state of the art.
Numerical discretization is only present in numerical simulation, and here just by adding brute force you can reduce is as much as you can. Current state of the art is not allowing to reduce it to 0.
Turbulence modelling is the key for numerical simulation. Current models used in the industry are not able to properly model some extreme areas (high separated flows for instance) and the current most promising techniques (DNS, LES) are highly demanding in computational time, increase constrains in the discretization area and, more importantly, have not been proved properly on industrial relevant cases.
You can see that current numerical simulation is ahead to wind tunnel (hardware simulation) in the area of geometry and external conditions, but mainly weak on the area of turbulence modelling.
Current situation is that wind tunnel has been replaced by numerical simulation in the area where turbulence modelling is not highly demanded by accuracy demands or physics complexity.
However, for being able to model some part of the physics the industry can only go hardware testing (wind tunnel).
Going more precisely to your question, numerical modelling will replace hardware testing when is able to solve the gap in turbulence modelling. The current situation as I described above, there are some approached promising to the future, but we will be able to evaluate if those approaches (DNS, LES) are able to close the gap when an application of a full aircraft is made. Currently, that is foreseen in 20 years time (I don't have reference sorry...), my point of view is that 20 years is optimistic.
So answering your question is, we can answer the question yet, we need 20 years further to be able to answer you.
Finally I would like to make a point in terms of costing. It is clear than creating a model in hardware testing (a phyisical model) is much expensive than a model in numerical testing (a mesh). However, a model in a wind tunnel has a huge productivity once in a wind tunnel (a new simulation point in less than a second), simulation takes hours once the model is ready. However, numerical simulation is reducing each year the cost of its model much more faster than wind tunnel. Still for some applications with high productivity of points needed, wind tunnel is far cheaper.