I don’t think this is true. There are many physical ststems for which we know the underlying physics very well, but the equations can’t be simply solved, and numerical simulation is more costly than just building the damn thing and testing it. Wing lift under turbulent conditions is one of those things. So we use wind tunnels. Not because we don’t understand lift—we do—but because it’s just easier.
This is getting less and less true with each generation of supercomputers though.
EDIT: There is perhaps a better way of explaining it for this crowd though. To use numerical modeling to predict performance is to take a physical problem and turn it into a computational problem. And while engineers understand physical systems pretty damn well, us computer scientists have largely failed at the objective of making software systems with hard reliability guarantees. You can write a fluid dynamics simulation to test your new wing design, but how do you know that the simulation does what you think it does? Even if the code has been tested before, how do you know you're not now hitting some sort of edge case?
At the end of the day, you have to build the damn thing to test it. Numerical simulation are used more and more these days as the codes are refined, computers get more powerful, and engineers have more trust in their capabilities. But traditionally, and still a lot of the time, they build prototypes and test in wind tunnels because reality never fails to model physics accurately.