CFD and AI

[mwillems]

Hey. I was thinking about areas of the car that feel like they may be more iterative in design than others due in part to the size and scope of airflow coming from them, such as the floor wing. We know that AI is used in the strategy sphere, but I was curious to understand how AI is used in the design of an F1 car and whether, given the complexity of CFD, AI at a point that is able to help push designs further along in their development so that the CFD and tunnel resources are focussed on more developed solutions than at present.

I had a quick google and I discovered that Williams are using technologies provided by Neural Concept which does exactly this.

It seems that their AI technology allows F1 teams to rapidly iterate and optimise vehicle designs by predicting aerodynamic performance. Williams, from what I can see so far in quotes, are cagey about how this is used and will only say that it is helping them find time and aid their development process. I wonder what constraints might sit around the type of CFD "problems" that might be put to this AI model, for instance how much interaction it has with other airflows, how complex are the airflows etc

Does anyone have any knowledge themselves of AI in F1 design, beyond what we know of this company, Neural Concepts? In itself it seems like a potential AI "Newey". Their product is called Neural Concept Shape and strictly speaking it isn't AI, it is Machine Learning which is a component of AI alongside NLP and the LLM. But, beyond that pedantry, it offers the same kind of benefits, it would seem.

https://techcrunch.com/2024/04/14/how-n ... rmula-one/

[noname]

You may want to look for information about using machine learning in scientific/engineering computing.
It is not a new thing, I am using it for about a decade.

[mwillems]

You may want to look for information about using machine learning in scientific/engineering computing.
It is not a new thing, I am using it for about a decade.

I'm specifically interested in knowledge of how it is used in cutting edge F1 design where the complexity dwarfs many fields.

I use ML and AI myself and am aware of its use for many applications. But linking in to F1 CFD or FFD processes is a different beast, especially given the need for accurate and usable results first time, to ensure the cost / benefit ratio under the cap and development restrictions work.

The Neural concepts solution is significantly expensive, for instance, and designed for Aero. I'd imagine any AI or ML tool supporting F1 is very much dedicated to Pre CFD analysis and not just any ML tool will satisfy this to the degree that F1 requires.

I'm also interested in understanding how accurate these solutions might be and to what degree they may be undermining development caps. Given the rate at which as is developing with Nvidia and the like releasing much more powerful AI chips, how long before you can accurately CFD model in AI?

One would imagine that CFD tools are also already looking to see where AI can be directly integrated to reduce compute time like Fast Fluid Dynamics already has on some areas.

[Greg Locock]

" complexity dwarfs many fields" and is about the same as any other complex engineering field. Silly fanboi nonsense.

[mwillems]

" complexity dwarfs many fields" and is about the same as any other complex engineering field. Silly fanboi nonsense.

You don't need to be rude. In terms of engineering design vs CFD in F1, why is it that so much compute time is required for CFD if it isn't so simple in terms of the calculations?

[Greg Locock]

I didn't say CFD was simple. I challenged the absurd assumption that F1 (making two cars go round a track for few 100 km) is the pinnacle of complexity in engineering.

[mwillems]

I didn't say CFD was simple. I challenged the absurd assumption that F1 (making two cars go round a track for few 100 km) is the pinnacle of complexity in engineering.

But I didn't say that either

I didn't mention anything about making a car go round a track, I talked about the level of detail used within CFD in F1, the number of interacting airflows, the detail these teams go to and the money and people they put at these designs is pretty substantial, more than many applications. It is approaching Aerospace levels of design, in fact Mclaren used to have an association with BAE. Except it is done within much tighter constraints of cost and time. The fact that CFD takes days and such huge computes to generate a model makes it fairly clear that it isn't something that ML can just pick up and therte is a big journey to improve the accuracy of any ML predictions or even integrate it directly within the F1 CFD process.

Let's try this another way, what can you actually explain about the use of AI or ML in the design processes of the F1 cars and how closely it is connected to the CFD process?

[Greg Locock]

well, gradient descent (Cauchy 1847) has been part of engineering since about 1950 (Haskell et al). It (and the general field of multivariate optimisation)was co-opted by the AI boys during one of the AI winters, but I can't say it really feels like AI to me.

[dialtone]

well, gradient descent (Cauchy 1847) has been part of engineering since about 1950 (Haskell et al). It (and the general field of multivariate optimisation)was co-opted by the AI boys during one of the AI winters, but I can't say it really feels like AI to me.

Ultimately it is though, that’s just how you find the weights.

It allows you to throw a bunch of features in the model and the optimization figures out their relationship. DL also uses the same mechanism to back propagate the weights.

As all optimization problems it’s very nice to minimize the error in your prediction. Incidentally it’s also incredibly scalable as you can do stuff like Parallel SGD to distribute it like Google (and others, but Google wrote the paper).

Anyway I’m sure you know this as you quoted the paper properly .