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).
The most reliable part of the model is the part covered by flight test data, where a pretty standard set of dynamic/kinematics equations is allied to a fairly standard set of aerodynamic coefficients and derivatives (CL-alpha, Cn-rudder, and so on) all defined as tables of data points as a function of the significant variables (so CL may be defined as a multi-linear function of alpha, Mach, flap position, etc.).
The data are derived from analysis of the data collected in flight test, and then confirmed by "matching" the model output to flight test output with the same set of inputs. (Pilot put in 3 deg elevator, aircraft pitched 2 deg/sec, sim model pitches 1.9 deg/sec, close enough, next data point ...)
Where no flight test data are available wind tunnel data may be used to extrapolate or interpolate from the flight data. Alternatively data for a similar aircraft may be used to guide extrapolation trends.
Only at the very start of a programme might CFD-based data be used, but even then the tendency is to start from a known model of a similar configuration and apply deltas, rather than start from scratch.
The approach of attempting to solve the flow conditions in real time is not used in industry. (Sweeping statement, but I've never heard of anyone tying that; it makes no real sense compared to doing the analysis offline, and would be near-impossible to obtain a sim qualification too)