I've asked around, and it seems all flight training programs use a combination of flight simulators and aircraft flights to train pilots. This works well for most pilots who fly airliners which don't spend much time in turbulent air or close to obstacles, and the flight simulator aerodynamics are very close to what is experienced in actual flight.

My question is where do the flight simulator aerodynamics models breakdown? I was wondering this in relation to rough air flying skills, where the aircraft or helicopter flow field interacts with surrounding environment (like inside a convective storm or near obstacles like trees and buildings) I would imagine these are situations that are hard to model since it depends on accurately modelling the interactions, which are difficult to run quickly enough for running a flight simulator. Fixing this by only doing flight training seems dangerous, especially for newbie pilots since the margin for error is very slim.

Personally, even though its more expensive than a flight simulator, a first person view drone would be a good intermediate step between the flight simulator and full scale training, with the response characteristics and control interface being the same as the full scale aircraft or helicopter.

That said, I don't know where the flight simulator aerodynamics begin to diverge from what is actually is experienced in real life, so the drone might be completely unnecessary. I'd love to see how this problem is resolved now.

  • 3
    $\begingroup$ That training drone would need to be full scale to have the same flight characteristics. So it would end up as the actual aircraft minus pilot plus a lot of equipment to make it remotely controllable and return comprehensive signals about attitude and motion. Which should all be fed back to what would essentially be a simulator. That said, where should it fly to not be a hazard for people on the ground? $\endgroup$ Sep 12, 2015 at 7:28
  • $\begingroup$ @RobVermeulen -- well, the Edwards range is a good place to start if you want to do funky test flying without worrying about hurting things on the ground.... $\endgroup$ Sep 14, 2015 at 2:45
  • 2
    $\begingroup$ You're right @UnrecognizedFallingObject. Didn't think that big, living in The Netherlands where anything taking off will soon appear over somebody's roof. Even if that roof is below sea level. $\endgroup$ Sep 15, 2015 at 7:22

3 Answers 3


Size matters here. Big aircraft have more inertia and take much longer to respond, but can equal out small-scale turbulence better. Your drone idea for training will not be representative of the big airplane at all.

Turbulence can be modeled quite faithfully - you just need to fly once through rough air, collect all the data and replay it in the simulator. The changes due to different pilot responses can be added on top without much loss in accuracy against a real flight.

Where models for airliner crew training break down is when the linear range of aerodynamics is left. Their post-stall and spin behavior is most likely not suitable for training, but that is not what they are designed for. Military aircraft simulators can even model the post-stall regime well, but need a lot of aerodynamic data, and most of that data is from wind tunnel models and not from real flight test.

What the simulation does is to plug the actual values of angle of attack, angle of sideslip, airspeed, power level angle and control deflections into matrices of coefficients and calculate the result from linear addition. That works well as long as the forces vary linearly of the flow angles, and correction factors will give you still good agreement with reality well into a stall. The coefficients are a mixture of CFD, wind tunnel data and flight test data. Gust loads can be modeled with Markov matrices which produce a realistic distribution of gust intensities, so even external, stochastic factors can be modeled faithfully.

I have less knowledge of helicopter simulators, and I would assume that flying near tall buildings can only be modeled in general terms, but not well enough to train a pilot for a specific location and weather condition.


There are two aspects to simulation error: the aerodynamic model error and less commonly known, integration algorithm numerical error.

Aerodynamic models tend to break down in the transonic and post stall regime of flight for two reasons: lack of data due to safety or cost, and poor predictability created by non-linear response in turbulent flow/shockwave formation. That is, the risk to reward would prohibit the conduct of full post stall(spin) test program on a wide body airliner. There is no justifiable reason to do so in relation to the aircraft's role or certification requirements. Contrast this to a fighter jet that might require this.

Buffet, ground effect, external turbulence, and wake turbulence can all be simulated with reasonable accuracy according to the requirements and complexity of the sim, which is dictated by the aircraft role. Note that buffet can occur prestall. Despite the difficulties in predicting post stall handling qualities using tools such a CFD analysis, it is possible to collect reasonable data in relation to handling characteristics with detailed and thorough flight test.

Simulators are usually hamstrung by model complexity and realtime computing power. For that reason, integration algorithms are chosen to best suit the simulation application. Airliner sims tend to use forward predictive algorithms such as Adams-Bashforth-Moulton predictors, as the flight envelope is limited to regions where strong linearity exists. Conversely this algorithm would induce significant errors even if reasonable aerodynamic data existed in post stall, spinning flight. Other algorithms do exist and are better suited for this application. Brute force may also be an option.

Using a drone could be beneficial, however it would be difficult to accurately mimic moments of inertia characteristics, and post stall handling due to Reynolds effects. Ok for generic simulation, but not ok when replicating larger scaled aircraft.

  • $\begingroup$ Would it still be a problem using euler integrators while running the simulation @ 3000 Hz? That rate seems to be no problem for an average laptop. $\endgroup$
    – Koyovis
    May 1, 2017 at 13:51
  • $\begingroup$ Simple Euler integration numerical methods can be a solution, but the problem depends of the complexity/resolution of calculations, which can only be done in serial, not parallel. My experience with an average lap top is that it can bog down quite quickly such that the errors increase significantly. $\endgroup$ Aug 3, 2017 at 21:00

Full motion simulators have no problems reproducing turbulent flight effects: they don't need to compute turbulence in real time like in CFD, they just replay one of several pre-recorded wind gust variants, and the reaction of the simulator flight model on this pre-recording is just a matter of mechanical physics modelling.

That is - as long as variables such as angle of attack stay within normal flying bounds, and do not enter fully established stall territory. That needs to be modelled in a different way, indeed like Peter says from wind tunnel data, combined (unfortunately) with data from flight recorders of crashed aircraft. This links to a site describing full stall simulation addition, certified by FAA. (I've long hesitated to link to it due to it being a commercial product and I've done business with the manufacturer).

Some confusion is created by Xplane using real-time laminar CFD. Turbulent real-time CFD has all kinds of challenges, amongst others the stability of the real-time loop like @busdrivingtupperware mentions. Flight training simulators don't use real-time CFD for computing flight dynamics, because computers are not yet powerful enough. Only a matter of time: code that used to run at 30 Hz on a super duper computer 25 years ago, now runs at 3000 Hz on a laptop. And model stability is usually not a problem at high iteration rates.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .