Is using design aids like AI and Machine Learning legal?

[ringo]

Can aero engineers and even chassis engineers use AI to refine and create surfaces, and if so how does this impact fairness and legality in F-1?
Forgive me if the technical regulations speak to this already.
But.. Having looked at the redbull floor and some of its other components, one could imagine that some of the surfaces were not designed by man. They may have been optimized by super computers and IA.
In the era of budget cap, an AI adapted for manufacturing and design can be developed outside of the sport by a sister company and simply sold to an F-1 as CAD software.

How will AI use and unmanned design be regulated in F-1, and are teams already using and benefiting from AI?

[Zynerji]

Careful. My thread on this didn't go well. What do they do with the CFD computers when they runout of allowance? Maybe they generative AI from the CFD output until the next allocation allotment?

[Greg Locock]

If topology optimisation is AI then no wonder ChatGPT is such a nonsense. Start with a big block of material. Load it up. Get rid of all the bits that aren't stressed. Rinse and repeat.

[Zynerji]

If topology optimisation is AI then no wonder ChatGPT is such a nonsense. Start with a big block of material. Load it up. Get rid of all the bits that aren't stressed. Rinse and repeat.

For FEA, I agree. For aero surfaces, I'd expect a point cloud that gets eroded after the initial expectations are set.

[Stu]

The limitation on CFD runs should put a lid on the use of optimisation software for aero surfaces (each optimisation run would count towards the allowance).

However…
There has been a suggestion (by Sam Collins in this weeks preview show), that cooling flows are not included within the run count - and this would seem to be an ideal use for such software, both in duct design and cooler design).

[Just_a_fan]

How would AI design aero components without running some sort of CFD implementation?

[n_anirudh]

How would AI design aero components without running some sort of CFD implementation?

It can predict flow fields on various 2D airfoils etc (which can be extended to 3D by CNN), so faster prediction based on previous "learnt" flowfields can be used for small design changes. Teams can utilise this from their vast CFD database/runs and train their AI codes and then use the learned weights to assess newer designs. Its mostly about uncertainty qualification and tuning their AI methods.

https://www.sciencedirect.com/science/a ... 58478#fig1

[chrisc90]

Ask the AI to run the CFD for you and report back. Use somebody else’s computer.

[Just_a_fan]

Ask the AI to run the CFD for you and report back. Use somebody else’s computer.

That would fall foul of the rules already, I think. Stuff done elsewhere has to allocated against your allowance. If it's done "on the sly" then it's simply cheating.

[chrisc90]

Ask the AI to run the CFD for you and report back. Use somebody else’s computer.

That would fall foul of the rules already, I think. Stuff done elsewhere has to allocated against your allowance. If it's done "on the sly" then it's simply cheating.

Scary question is… how long will it be before AI is able to do that sort of thing and return a good result.

If you ask AI to make you a front wing with a little bit of outflow from the edges, plenty of load etc etc and a design has to fit within the volumes of these areas. How long before it comes back with a. Design worthy? How does any time get allocated? If AI does say 15 runs in the background on a design and presents the best one to the person asking… how does F1 interpret the designs in the background ?

I doubt AInis good enough yet though however

[Just_a_fan]

How would AI design aero components without running some sort of CFD implementation?

It can predict flow fields on various 2D airfoils etc (which can be extended to 3D by CNN), so faster prediction based on previous "learnt" flowfields can be used for small design changes. Teams can utilise this from their vast CFD database/runs and train their AI codes and then use the learned weights to assess newer designs. Its mostly about uncertainty qualification and tuning their AI methods.

https://www.sciencedirect.com/science/a ... 58478#fig1

Skimmed it, interesting paper. Worked on a very small mesh c.1500x1500 and very low Reynolds numbers c. 10^-4 to 10^-2 and took 800 CFD runs to provide the learning dataset. And got c.3.5% error

I wonder how translatable that is to something of the size of a car component. 3.5% error would be way too big for F1 aero work, I'd have thought.

AI feels a bit like graphene from a few years ago. Yes, there are things it's very useful for, but it's not the universal tool that the media like to portray it.

[Just_a_fan]

Ask the AI to run the CFD for you and report back. Use somebody else’s computer.

That would fall foul of the rules already, I think. Stuff done elsewhere has to allocated against your allowance. If it's done "on the sly" then it's simply cheating.

Scary question is… how long will it be before AI is able to do that sort of thing and return a good result.

If you ask AI to make you a front wing with a little bit of outflow from the edges, plenty of load etc etc and a design has to fit within the volumes of these areas. How long before it comes back with a. Design worthy? How does any time get allocated? If AI does say 15 runs in the background on a design and presents the best one to the person asking… how does F1 interpret the designs in the background ?

I doubt AInis good enough yet though however

It's not really AI though, it's just a bot running CFD on remote resources.

For AI to come up with a design without solving NS equations (I.e. not doing some form of CFD), it would need to be trained on a huge dataset of existing wings. Most of the wings relevant to an F1 team are F1 wings - and those are proprietary and so wouldn't be available for the learning dataset. So a team would just be using its own data to train the system. That's a huge limitation.

[chrisc90]

Its been done with race strategy so probably not too much harder to do car design.

It would ruin the sport though. Where is the fun in watching computer generated designs race round a track.

AI would have a melt down if Adrian Newey tried to teach it by looking at a notepad + pencil and a sketch of something on his drawing board

[JordanMugen]

Where is the fun in watching computer generated designs race round a track.

What's the difference between an engineer iteratively generating a part by looking at pressure or stress distributions (aero or structure) in a simulation and tweaking to suit, compared to the computer iteratively generating the part as well as doing the simulations?

[Just_a_fan]

Its been done with race strategy so probably not too much harder to do car design.

Doing an entire car is way more involved than strategy - Monte Carlo simulation of strategy is one thing, using it to design a car is not something one is going to do. At least not yet, anyway.