search

convexengineering / SPaircraft

Models for commercial aircraft design
http://spaircraft.readthedocs.org
25 stars 17 forks source link

all D8 models have relaxed constants at the final GP iteration #52

Closed mayork closed 7 years ago

mayork commented 7 years ago

@1ozturkbe this is not due to the subs, I checked. Any idea what is going on? Did you change the root moment constraints on the D8 and 737 type models?

mayork commented 7 years ago

could it be the buoyancy stuff?

mayork commented 7 years ago

could this do with the new wing sizing for landing load? 777 isn't solving without relaxed constants either and Drela said that he doesn't size the wing for landing and I don't think we should. In reality you could just impose a max landing fuel to solve the landing load issue and size the wing for aero loads...no sense in carrying around extra structure for no reason

1ozturkbe commented 7 years ago

@mayork Pretty certain it is not the buoyancy stuff, because this was occurring way in advance of adding that model. Definitely not a consequence of the wing landing model either; I will likely remove that today because it is not actually present in TASOPT. (This is intriguing to me because the 73 is designed to land at max fuel, since it has no jettison system.)

mayork commented 7 years ago

I think hunting this down needs to be top priority...it's gotta be one of the differences between 737 and D8

1ozturkbe commented 7 years ago

After some testing, I believe that it is something to do with the solver. If I solve for the D82 with a reltol=0.01, the model solves (output well behaved, no relaxed constants), but at higher reltol's the model fails. @bqpd can you help us decipher what this may mean? I'm pretty certain at this point that the model is well formulated, and that all of the necessary substitutions are in place.

1ozturkbe commented 7 years ago

Furthermore, I realize this happens when the initial objective value of the solver is much lower than the previous optimum, but rises rapidly during the solve to reach the previous optimal value. Perhaps some odd solver dynamics?

mayork commented 7 years ago

great work Berk this is certainly interesting...I hadn't even thought to look at the solver. I might try solving with cvxopt later (it'll take forever but might be interesting). For now, should we just set the higher reltol value and push on?

1ozturkbe commented 7 years ago

I don't see any reason not to! Plus, lower reltol speeds up the solution speed, and means that we do not do a large number of iterations hopping around the optimal value like we used to.

mayork commented 7 years ago

closing this issue...tends to be caused by unrealistic subs, particularly in the CG