Is using design aids like AI and Machine Learning legal?

[AR3-GP]

Teams are already using adjoint CFD methods. This is already a computer aided approach to shape optimization by way of calculating gradients of the various constitutive equations as a function of the boundary conditions.

I think AI is a long way off being meaningful here. Forget about machine learning. These local adjoint methods are a much more efficient way to optimize shapes as they are rooted in calculations of the actual physical quantities in question. Machine learning and AI are attempts to not use physics to derive aero shapes. This is massively inferior.

[n_anirudh]

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?

Inverse mapping of Cp exists, and if a desired pressure distribution exists, one can generate an airfoil shape based on it.

[n_anirudh]

Teams are already using adjoint CFD methods. This is already a computer aided approach to shape optimization by way of calculating gradients of the various constitutive equations as a function of the boundary conditions.

I think AI is a long way off being meaningful here. Forget about machine learning. These local adjoint methods are a much more efficient way to optimize shapes as they are rooted in calculations of the actual physical quantities in question. Machine learning and AI are attempts to not use physics to derive aero shapes. This is massively inferior.

As with any tech, they get better over time. Just look at computer vision and implementation of instance segmentation
https://ai.meta.com/blog/segment-anythi ... mentation/. This will be the norm 3-4 years down the line

[AR3-GP]

Teams are already using adjoint CFD methods. This is already a computer aided approach to shape optimization by way of calculating gradients of the various constitutive equations as a function of the boundary conditions.

I think AI is a long way off being meaningful here. Forget about machine learning. These local adjoint methods are a much more efficient way to optimize shapes as they are rooted in calculations of the actual physical quantities in question. Machine learning and AI are attempts to not use physics to derive aero shapes. This is massively inferior.

As with any tech, they get better over time. Just look at computer vision and implementation of instance segmentation
https://ai.meta.com/blog/segment-anythi ... mentation/. This will be the norm 3-4 years down the line

There's a difference in the kind of problem which is being solved here. Identifying objects is a different task.

Your link does suggest a potential avenue for AI assisted interpretation of flow structures from CFD and PIV. AI could also be used to pour over data and look for trends which are not obvious. For example, AI may be able to provide pecking orders for the second half of the year based on performance trends in the first half of the year. You could replace a human job with that.. AI could also assist you in identifying what to study, but you will still need to do a CFD run or a windtunnel run to tell you that the car is better. That part can't be done without physics or a calculation of some kind.

AI and ML will improve over time in solving the problems that they are good at solving. So you have to know when to use it.

[n_anirudh]

https://www.sciencedirect.com/science/a ... via%3Dihub
^^These methods have similar potential.

The previous link is for image segmentation. This can be used for parts inspection and to detect cracks and faults after each GP when cars are inspected. Alternately, the size of the vortex coming off a front wing/location based on CFD imagery. Agreed its not physics based, but there is PINN - https://arxiv.org/pdf/2105.09506.pdf

The operative word is to assist. There will be a manual or semi automated validation of final designs. Think of it as CFD in the late 90s or early 2000's. CFD became an expert subject to something every undergrad studies. With AI and ML, and the compute power available, the adoption will be a lot faster now.

[Just_a_fan]

Is there a risk that using the generic term "AI" hides the actual techniques used? Perhaps there needs to be use of more definitive terms than just "AI".

Even when we talk about AI we usually mean Artificial Narrow Intelligence (ANI) - something taught using a huge dataset and which basically looks at that data and compare it to the question asked. It's not making leaps of understanding and it can do no more than what it is asked to do with the dataset it has. We'd need to move to Artificial General Intelligence (AGI) to have a system that can move outside of the dataset it has to make the cognitive leaps we want.

And the techniques being discussed here in various linked papers aren't "AI" in any way other than the general public's idea of AI as fed to them by the mass media.

[AR3-GP]

https://www.sciencedirect.com/science/a ... via%3Dihub
^^These methods have similar potential.

The previous link is for image segmentation. This can be used for parts inspection and to detect cracks and faults after each GP when cars are inspected. Alternately, the size of the vortex coming off a front wing/location based on CFD imagery. Agreed its not physics based, but there is PINN - https://arxiv.org/pdf/2105.09506.pdf

That's an outstanding idea. You could use AI to look at X-Ray and ultrasound imaging to detect cracks.

I think my main point (and the question of the OP), was using AR/ML as an aero design aid and whether that would be legal or not. I think that's a more provocative/controversial question.

[Just_a_fan]

The whole "AI does CFD" thing really needs us to arrive at an AGI that can "see the air" (in the way that people like to say Newey does). In effect it would be doing CFD in its head (difficult to come up with a suitable metaphor). At that point, one can start to see that the legality question is raised - is an AGI "seeing the air" a CFD process controlled by the resource rules? I think most sensible people would say yes it is, but this is F1...

[AR3-GP]

The whole "AI does CFD" thing really needs us to arrive at an AGI that can "see the air" (in the way that people like to say Newey does). In effect it would be doing CFD in its head (difficult to come up with a suitable metaphor). At that point, one can start to see that the legality question is raised - is an AGI "seeing the air" a CFD process controlled by the resource rules? I think most sensible people would say yes it is, but this is F1...

Any CFD at all, must be accounted for and I feel that's rather logical. The AI should merely act as an operator. The big advantage here is the budget cap. Why pay an employee to queue surfaces for CFD when a bot could do it for the cost of the CPU and electricity to run it.

I also wonder if there should not be some regulations surrounding automation...If the original goal of the cap was to in part stop teams hoarding employees like task rabbits, then what is the equivalent policing method when teams are just replacing people with computers for pennies on the dollar in "labor" cost.

Many teams had to downsize. The biggest teams had over 1000 employees before. They have downsized due to budget cap which made them closer in size to smaller teams, but if the big teams are just growing again by way of robots, automation, and AI, then they are basically still operating like 1000 person teams.

[maygun]

Teams are already using adjoint CFD methods. This is already a computer aided approach to shape optimization by way of calculating gradients of the various constitutive equations as a function of the boundary conditions.

I think AI is a long way off being meaningful here. Forget about machine learning. These local adjoint methods are a much more efficient way to optimize shapes as they are rooted in calculations of the actual physical quantities in question. Machine learning and AI are attempts to not use physics to derive aero shapes. This is massively inferior.

As with any tech, they get better over time. Just look at computer vision and implementation of instance segmentation
https://ai.meta.com/blog/segment-anythi ... mentation/. This will be the norm 3-4 years down the line

There's a difference in the kind of problem which is being solved here. Identifying objects is a different task.

Your link does suggest a potential avenue for AI assisted interpretation of flow structures from CFD and PIV. AI could also be used to pour over data and look for trends which are not obvious. For example, AI may be able to provide pecking orders for the second half of the year based on performance trends in the first half of the year. You could replace a human job with that.. AI could also assist you in identifying what to study, but you will still need to do a CFD run or a windtunnel run to tell you that the car is better. That part can't be done without physics or a calculation of some kind.

AI and ML will improve over time in solving the problems that they are good at solving. So you have to know when to use it.

In a similar sense, CFD also approximates real-world physic it is not real physics.
You can approximate physic also with AI/ML. With enough data, you can train a system that maps given 3D to CFD output.

For instance, weather prediction is generally based on numerical models similar to CFD (solving equations to approximate the physic of particles). But new ML/AI-based systems become superior, like this method: https://www.deepmind.com/blog/nowcastin ... ur-of-rain

[Just_a_fan]

You can approximate physic also with AI/ML. With enough data, you can train a system that maps given 3D to CFD output.

But the highlighted bit is the important bit, isn't it? It's getting a large enough dataset that's the issue for a team, because their own data, whilst plentiful, will be inherently limited by their own ability to generate meaningful output.

I doubt any team has a large enough dataset for ML to generate wing profiles in lieu of CFD.

[AR3-GP]

You can approximate physic also with AI/ML. With enough data, you can train a system that maps given 3D to CFD output.

But the highlighted bit is the important bit, isn't it? It's getting a large enough dataset that's the issue for a team, because their own data, whilst plentiful, will be inherently limited by their own ability to generate meaningful output.

I doubt any team has a large enough dataset for ML to generate wing profiles in lieu of CFD.

There are loopholes in the regulations for CFD/WT usage related to "methodology development". Make of that what you will...

[n_anirudh]

I see AI/ML as another tool that can assist in making the design process quicker. As with any tool/method, there are uncertainties and correlations need to be built, just as they calibrate their CFD with WT measurements and track runs.

I do think teams have sufficient data from 2D and 3D runs of their cars/parts. For instance, we may need about 1500 airfoils for training and that should be easily available to teams when they are iterating over a design process or from their optimisation runs IMO

[Just_a_fan]

Ultimately, the AI/ML thing is only saying "these are some things worth trying", they're not saying "build this, it's perfect", so the teams are still going to have use CFD and WT resources to assess the options chosen by the system. Whilst the system might allow them to consider a lot of profiles, I do wonder whether it will be any better than the intuition of an experienced aerodynamicist. These guys have got an idea about what should/won't work so they're already working with a limited set of options to trial.

The ML systems currently in use are generally interpolative - which means they can't be used to come up with some new design that isn't already contained within the known dataset. Extrapolating outside the dataset is much more difficult.

[maygun]

You can approximate physic also with AI/ML. With enough data, you can train a system that maps given 3D to CFD output.

But the highlighted bit is the important bit, isn't it? It's getting a large enough dataset that's the issue for a team, because their own data, whilst plentiful, will be inherently limited by their own ability to generate meaningful output.

I doubt any team has a large enough dataset for ML to generate wing profiles in lieu of CFD.

I would say every team has enough data probably. In 3D, even one pair instance contains very valuable information, if you can define a set of good regularizations, you can reduce the space of possible solutions, hence the requirement for more data becomes less of an issue.

I would make an arrogant statement here, with 4-5 good AI engineers, and good structured historical data of an average F1 team, you can build a system that generates designs that beat most of the current designers in one year.