Jump to content


Photo
- - - - -

Use of neural networks with CFD in F1 aero design.


  • Please log in to reply
4 replies to this topic

#1 TheWilliamzer

TheWilliamzer
  • Member

  • 1,208 posts
  • Joined: April 12

Posted 11 March 2019 - 21:33

This is one of the stream of ideas I had while in the toilet. I wondered if any team actually considers using computer-assisted aerodynamic design or am I just a dreamer.

 

I tried to think about how such a system might work and what is needed to achieve that, and I found these points:

 

1- the technical rules. Every bit of the aero of the F1 car is defined in the FIA tech regs. The minimum, the maximum size and the position of every element are defined and confined within margins. So the starting point in the programming process is to enter all these values and create the "area of creativity" where the magic will happens and all within the rules so no risk of illegal designs.

 

2- the neural network.

2-a: creation. in 3D space, an evolution based network generates random vertices and meshes within the "area of creativity" according to the set parameters, the operation is repeated many times until there's an important number of "cars" a.k.a models.

 

2-b: the scoring system. a CFD based system runs simulations on the generated "cars", and the ones with the best attributes (downforce vs drag) are kept and promoted to the next generation, with their dimensions and parameters dictating the margins of the "zone of creativity" to norrow it to optimal level.

 

the two steps above will repeat as much you like, and from now on, it's just letting the computers cook the recipes.

 

 

 

from time to time, some generations might include new "mutants" that might dramatically change the whole aero or just die in the few next generations. and the longer the network will be up, the better the results will be.

 

 

some challenges might arise, as the standards of F1 CFD require some very highly detailed models and very high particle density. and running one sim might need a CPU farm so let alone running tens, even hundreds of randomly generated subjects at once. but using the method for individual parts might be possible.

 

 

 

 

 

so that's me thinking in high voice and being dreamy. your comments are welcome!



Advertisement

#2 pdac

pdac
  • Member

  • 18,648 posts
  • Joined: February 10

Posted 11 March 2019 - 21:38

If it were viable, they would be doing it now.



#3 PayasYouRace

PayasYouRace
  • Racing Sims Forum Host

  • 52,426 posts
  • Joined: January 10

Posted 11 March 2019 - 21:50

Given that I was doing this on a basic level at university 10 years ago, I'm sure the F1 teams are all over it.



#4 skidmarkz

skidmarkz
  • Member

  • 63 posts
  • Joined: March 17

Posted 12 March 2019 - 16:44

If it were viable, they would be doing it now.

Generative design is viable and teams aren't doing it yet



#5 CPR

CPR
  • Member

  • 5,995 posts
  • Joined: February 10

Posted 12 March 2019 - 17:20

Don't forget that there's a limit to CFD usage in the regulations.

 

In other words, there's a large benefit to efficient usage of CFD. A genetic algorithm or similar doesn't strike me as being particularly efficient. I'm not saying there's no scope at all but it would probably be limited in scope.

 

Talking of the CFD limits, I believe it was on the total amount of floating point operations or something. When I saw this video a few months ago, I wondered if it was applicable to F1 CFD:

 

The above video is about how SpaceX used wavelet compression to create a sparse matrix and then run rocket engine simulation on a GPU. They also used it for simulating things like airflow around a capsule during re-entry.