Flight models on training sims, and used by manufacturers for engineering analysis, are inevitable data-based and in effect empirical models. Things like CFD are not used in the model (though may be used to extend the dataset to areas of the flight envelope where no other data exists).
How much of the flight envelope is usually covered by hard test data, and how much has to be simulated through extrapolation?
I've heard about simulators sometimes not being very good at extrapolating to flight regimes that weren't part of the tests. Wasn't there an accident involving extreme use of rudder controls or something some years ago that was traced in part to a difference between the way the sim behaved and the way the actual aircraft behaved.