Are F1 team using AI to gain advantage?

[Greg Locock]

I'm pretty sure it wouldn't because data capture is the hardest bit of expert systems.

[dialtone]

Generative AI is SLOOOOOOOOOWWWW and stupidly expensive for the volume needed here.

Yeah it could generate data but there are stochastic methods that are much faster and not that different ultimately, like setup a markov chain and go to town.

[maygun]

So onto the car design . Machine learning is what you would use for these very "constrained" and "guided" problems (regulations!) of a physical nature to help you make decisions. You could generate a fancy rear wing if you want with generative AI but that might be very out there, you have to test if it it works and you might have to test hundreds of these! but you can use machine learning to make the model of how to design a rear wing, and then choose the best design.

You can actually design a generative AI that makes performance predictions as well.

Let's say I want to design a Rear Wing, and my objective is to maximise the downforce, and they allocate my 10k CFD runs.

Using regulations, we can define the space of shapes for the rear wing using variables. Let's say 50 variables define the final shape of the rear wing. All of the variables are continuous with fixed ranges. Here, even if we opted to choose 3 random values (let's hope no team is doing a step-sweep search on multi-dim hyperparameter search ) for each variable it is nearly impossible to try all of the designs in CFD because we would end up 3^50 different configurations.

Here what we can do is, generate 9k random rear wing configuration (x_{i}), and measure the downforce of the rear wings (y_{i}) with CFD. Then we can learn (train using ml algorithms) a function f_{down_force} that takes a rear wing configuration and predicts the downforce value of that rear wing.

Then we can learn a new generative model that conditions the downforce level and produces a plausible rearwing that matches the downforce. Here we would learn a generative ml model g (z|y), where z is a random vector and y is the conditioning variable (downforce). This g function basically would learn to generate rear wings that match the conditioning downforce level.

The training of function g would be like this, generate a random rear wing (x_{j}) with downforce level y_{j}, first x_{j} should follow a similar distribution as x_{i}s and second the downforce level should match y_{j}. The downforce level of x_{j} would be measured using the f_{down_force} function that I learned in the previous step.

Then after the training of g is done, I would generate 1k wings with the maximum downforce values, and test these wings on the CFD/Wind tunnel.

[Hoffman900]

That isn’t AI though and people were doing that ten years ago under the guise of machine learning.

Furthermore, that only works if all iterations correlate to the real world. We know it’s not that simple. Because if there are 10k iterations it’s going to bias to the code and produce something that is only good in CFD.

[maygun]

That isn’t AI though and people were doing that ten years ago under the guise of machine learning.

Furthermore, that only works if all iterations correlate to the real world. We know it’s not that simple. Because if there are 10k iterations it’s going to bias to the code and produce something that is only good in CFD.

AI is machine learning.

In the last 10 years, there has been a huge development in generative models in machine learning, as well as 3D and graph networks, where you can process different types of input modalities.

There are other also huge amounts of works that deal with simulation/real-world gap problems, especially in the robotics domain within the RL algorithms. You can transfer a lot of the methods from there to solve or minimize the issues related to CFD real world gap.

[PlatinumZealot]

So onto the car design . Machine learning is what you would use for these very "constrained" and "guided" problems (regulations!) of a physical nature to help you make decisions. You could generate a fancy rear wing if you want with generative AI but that might be very out there, you have to test if it it works and you might have to test hundreds of these! but you can use machine learning to make the model of how to design a rear wing, and then choose the best design.

You can actually design a generative AI that makes performance predictions as well.

Let's say I want to design a Rear Wing, and my objective is to maximise the downforce, and they allocate my 10k CFD runs.

Using regulations, we can define the space of shapes for the rear wing using variables. Let's say 50 variables define the final shape of the rear wing. All of the variables are continuous with fixed ranges. Here, even if we opted to choose 3 random values (let's hope no team is doing a step-sweep search on multi-dim hyperparameter search ) for each variable it is nearly impossible to try all of the designs in CFD because we would end up 3^50 different configurations.

Here what we can do is, generate 9k random rear wing configuration (x_{i}), and measure the downforce of the rear wings (y_{i}) with CFD. Then we can learn (train using ml algorithms) a function f_{down_force} that takes a rear wing configuration and predicts the downforce value of that rear wing.

Then we can learn a new generative model that conditions the downforce level and produces a plausible rearwing that matches the downforce. Here we would learn a generative ml model g (z|y), where z is a random vector and y is the conditioning variable (downforce). This g function basically would learn to generate rear wings that match the conditioning downforce level.

The training of function g would be like this, generate a random rear wing (x_{j}) with downforce level y_{j}, first x_{j} should follow a similar distribution as x_{i}s and second the downforce level should match y_{j}. The downforce level of x_{j} would be measured using the f_{down_force} function that I learned in the previous step.

Then after the training of g is done, I would generate 1k wings with the maximum downforce values, and test these wings on the CFD/Wind tunnel.

Good. Though your example is very constrained and can already be achieved. Solidworks cando whatyou descibe in your first pargraph. I think you can be more couragous and let the AI create new holes, flaps etc wherever it wants instead of trying to fit into a mould.

Just like generative AI text, music, videos and images, even a generative AI wing will have to be tuned and optimized.. Maybe with machine learning again.

Also remember teraflops are limited and I think it applies to all computations. Generative AI is a heavy consumer when it comes to teraflops, memory, heat generations and data generation. So FIA will have to create new rules.

I think teams will be using this for strategies and team structure first.. (Scrap the suspension idea! I don't think it will be good at multiple parts machines yet! ).