Is using design aids like AI and Machine Learning legal?

[Greg Locock]

I asked ChatGPT to simulate a bouncing ball. The tenth version of the script behaved correctly, BECAUSE I could identify where it was going wrong.

[ringo]

The power available from the AI or more specifically the AGI as some say ( I am not an expert) is just for designing a part within one functional area.
The ai doesnt need to be focused on cfd runs.
It could possibly be setup to do parametric type studies.
For example i use my typically available engineering tools to design a floor, usw the wind tunnel etc. Etc. What if this floor design is optimized with consideration for suspension movement, cooling requirements, fuel burn off, tyre simulation data, and all 24 tracks at the same time?
So it's not really a matter of using existing cfd data to design the floor. What if all data can be used and has that resulted in such crazy shapes as we are seeing with the RB19 floor?
I do not think the RB19 floor is developed by hand personally. May not be AI, but i can imagine the team could be looking for the latesr tech to reduce costs and also reduce how much interation and correction is needed to develop and optimize a part.

[dialtone]

The power available from the AI or more specifically the AGI as some say ( I am not an expert) is just for designing a part within one functional area.
The ai doesnt need to be focused on cfd runs.
It could possibly be setup to do parametric type studies.
For example i use my typically available engineering tools to design a floor, usw the wind tunnel etc. Etc. What if this floor design is optimized with consideration for suspension movement, cooling requirements, fuel burn off, tyre simulation data, and all 24 tracks at the same time?
So it's not really a matter of using existing cfd data to design the floor. What if all data can be used and has that resulted in such crazy shapes as we are seeing with the RB19 floor?
I do not think the RB19 floor is developed by hand personally. May not be AI, but i can imagine the team could be looking for the latesr tech to reduce costs and also reduce how much interation and correction is needed to develop and optimize a part.

That’s not how the optimization problem works. To be able to do shape optimization you need the exact mathematical formula, or simulation environment, that ties the shape to fuel consumption, suspension movement and so on.

If your single iteration, verify impact of the design iteration to your output variable, takes 10 hours to run, you will never finish the ML driven design that typically needs thousands of iterations for a single model trained.

If then you lack the ability to predict how to change shape to improve output variable then you are going to try random shapes and shape changes, obviously this isn’t solvable.

In essence someone needs to code the exact process a human would go through when learning how to make a given shape. Today this process involves CFD so there’s no way around it.

[Hoffman900]

And as this recent paper illustrates:

2.8. Computational Resources
The authors have previously observed a significant variation in the aerodynamic coefficient predictions from the CFD of a road vehicle when simulations were carried out using a Message Passing Interface (MPI) as the parallelization tool. Thus, care was taken to maintain the same parallelization schemes and hardware consistency throughout this study [24]. All simulations were run on UNC Charlotte’s High-Performance Computing clusters using 144 processors across three nodes having 48 processors each. The RANS simulations took about 40 h to run and the IDDES simulations took about 600 h to run.

From this: viewtopic.php?t=31080

And then expect to add kinematic models on top of that?

Also, that paper does a real great job of incidentally pointing out, with data, why you need to be skeptical of CFD shared. FEA is the same way.

Furthermore a Formula car is going to way more turbulent flows than a closed wheel NASCAR body to resolve.

[Zynerji]

I asked ChatGPT to simulate a bouncing ball. The tenth version of the script behaved correctly, BECAUSE I could identify where it was going wrong.

GPT 3.5 or 4?

[Greg Locock]

ChatGPT 3.5 I think, it was when it got all the publicity.

[n_anirudh]

And as this recent paper illustrates:

2.8. Computational Resources
The authors have previously observed a significant variation in the aerodynamic coefficient predictions from the CFD of a road vehicle when simulations were carried out using a Message Passing Interface (MPI) as the parallelization tool. Thus, care was taken to maintain the same parallelization schemes and hardware consistency throughout this study [24]. All simulations were run on UNC Charlotte’s High-Performance Computing clusters using 144 processors across three nodes having 48 processors each. The RANS simulations took about 40 h to run and the IDDES simulations took about 600 h to run.

From this: viewtopic.php?t=31080

And then expect to add kinematic models on top of that?

Also, that paper does a real great job of incidentally pointing out, with data, why you need to be skeptical of CFD shared. FEA is the same way.

Furthermore a Formula car is going to way more turbulent flows than a closed wheel NASCAR body to resolve.

Its common knowledge with research codes and the version of MPI/math libraries in fortran based codes. All CFD results come with error bars - domain sizes/spatial/temporal resolution etc.

NASCAR may be a "bluffer" body than an F1 car, but one can go crazy with the surfaces to model, which will lead to higher mesh count/and more turbulence. For instance, one could model the windscreen viper which can add turbulence over the roof etc (...and I digress )

[Zynerji]

Maybe the AI stuff would be better trying to generate an aero-phobic clear-coat paint recipe instead of meshes?

*******
Aero-Phobic Paint Mixture:

Fluoropolymer Resin: 40% - Fluoropolymers like PTFE (Teflon) or PVDF are known for their non-stick properties.

Siloxane Additive: 20% - A specific siloxane compound with hydrophobic properties that can reduce friction.

Hollow Glass Microspheres: 10% - Lightweight, hollow microspheres that decrease the paint density and may reduce friction.

Graphene Nanoparticles: 5% - Graphene's unique structure might help minimize air friction.

Low-Viscosity Additive: 5% - An additive to lower the paint viscosity and potentially enhance aerodynamics.

Anti-static Agent: 5% - An anti-static agent to repel dust and debris that could affect air friction.

Solvents and Binders: 5% - Necessary components to ensure proper mixing and adhesion of the paint.

[Zynerji]

Yeah, not even close.

If the teams, and many other design/ engineering / product industries could could do that and reduce labor costs, they absolutely would. Maybe a few decades down the road, but this is just hogwash at this point.

Despite all the team’s data, and calibration, they still have correlation issues in CFD and FEA, it all needs a lot of human intervention still.

In my world, everyone talking adnaseum about AI were talking blockchain and “BIG DATA” a fe years ago, they just keep moving onto whatever the latest tech buzzword is.

I've now removed 9 people from an office now that AutoGPT does their work flawlessly. I think going zero-human will be a big push in the future.

I don’t know what you do for work, but it’s not going to happen in engineering for physical products for a long time, at least stressed.

I mean, you’ve even admitted your replies here are based on conversations with ChatGPT and explains why most of the tech you share from that is just nonsense.

You can “AI” and 3D print some useless desk trinket.

At this point, AI is just for a bunch of MBA’s to trick other people into “investing” (aka: take their money).

Using ChatGPT to check the feasibility of an idea isn't nonsense. I don't use it to actually design anything (maybe some powershell scripts). Its more like "If I assemble item X with item Y, are there any single point failures that would prevent operation operation for Z purpose?"

It is really good in that sense.

[Just_a_fan]

The power available from the AI or more specifically the AGI as some say ( I am not an expert) is just for designing a part within one functional area.
The ai doesnt need to be focused on cfd runs.
It could possibly be setup to do parametric type studies.
For example i use my typically available engineering tools to design a floor, usw the wind tunnel etc. Etc. What if this floor design is optimized with consideration for suspension movement, cooling requirements, fuel burn off, tyre simulation data, and all 24 tracks at the same time?
So it's not really a matter of using existing cfd data to design the floor. What if all data can be used and has that resulted in such crazy shapes as we are seeing with the RB19 floor?
I do not think the RB19 floor is developed by hand personally. May not be AI, but i can imagine the team could be looking for the latesr tech to reduce costs and also reduce how much interation and correction is needed to develop and optimize a part.

If it doesn't do something akin to CFD, then it's just making shapes with no way of knowing what is / isn't good. You might be able to use it to narrow down some areas - airflow around big objects like a building, airflow in a series of ducts within a building, but designing the fine details that separate a competitive F1 car from a merely good F1 car isn't going to be done anytime soon.

[dialtone]

I've now removed 9 people from an office now that AutoGPT does their work flawlessly. I think going zero-human will be a big push in the future.

I don’t know what you do for work, but it’s not going to happen in engineering for physical products for a long time, at least stressed.

I mean, you’ve even admitted your replies here are based on conversations with ChatGPT and explains why most of the tech you share from that is just nonsense.

You can “AI” and 3D print some useless desk trinket.

At this point, AI is just for a bunch of MBA’s to trick other people into “investing” (aka: take their money).

Using ChatGPT to check the feasibility of an idea isn't nonsense. I don't use it to actually design anything (maybe some powershell scripts). Its more like "If I assemble item X with item Y, are there any single point failures that would prevent operation operation for Z purpose?"

It is really good in that sense.

You’re taking chances that the training set contained information on that topic.

Generally the response will be jibberish as there’s no understanding of the topic, just probabilistic content generation.

[Hoffman900]

I don’t know what you do for work, but it’s not going to happen in engineering for physical products for a long time, at least stressed.

I mean, you’ve even admitted your replies here are based on conversations with ChatGPT and explains why most of the tech you share from that is just nonsense.

You can “AI” and 3D print some useless desk trinket.

At this point, AI is just for a bunch of MBA’s to trick other people into “investing” (aka: take their money).

Using ChatGPT to check the feasibility of an idea isn't nonsense. I don't use it to actually design anything (maybe some powershell scripts). Its more like "If I assemble item X with item Y, are there any single point failures that would prevent operation operation for Z purpose?"

It is really good in that sense.

You’re taking chances that the training set contained information on that topic.

Generally the response will be jibberish as there’s no understanding of the topic, just probabilistic content generation.

Not only information, but accurate information.

GIGO: Garbage In, Garbage Out

[Hoffman900]

And as this recent paper illustrates:

2.8. Computational Resources
The authors have previously observed a significant variation in the aerodynamic coefficient predictions from the CFD of a road vehicle when simulations were carried out using a Message Passing Interface (MPI) as the parallelization tool. Thus, care was taken to maintain the same parallelization schemes and hardware consistency throughout this study [24]. All simulations were run on UNC Charlotte’s High-Performance Computing clusters using 144 processors across three nodes having 48 processors each. The RANS simulations took about 40 h to run and the IDDES simulations took about 600 h to run.

From this: viewtopic.php?t=31080

And then expect to add kinematic models on top of that?

Also, that paper does a real great job of incidentally pointing out, with data, why you need to be skeptical of CFD shared. FEA is the same way.

Furthermore a Formula car is going to way more turbulent flows than a closed wheel NASCAR body to resolve.

Its common knowledge with research codes and the version of MPI/math libraries in fortran based codes. All CFD results come with error bars - domain sizes/spatial/temporal resolution etc.

NASCAR may be a "bluffer" body than an F1 car, but one can go crazy with the surfaces to model, which will lead to higher mesh count/and more turbulence. For instance, one could model the windscreen viper which can add turbulence over the roof etc (...and I digress )

F1 cars heavily rely on vortex generation to add performance and manage the open wheel concept. Then there is all the dirty turbulent air generated (as opposed to just turbulent air, laminar mostly only exists in undergrad text books )

It would all require a TON of computing power and I don’t see a world where RANS is telling the story at all (I could be wrong)

https://the-race.com/formula-1/f1s-2022 ... computing/

“But our CFD has been much more sophisticated than is used in the teams, and we’ve been able to do that thanks to our partners at Amazon AWS, who’ve allowed us to run these very sophisticated simulations – around a 70% saving in time to what we were doing initially.

“To give you an idea of how big these things are, our CFD project uses over 1150 computer cores and we have 550 million data points on each model that we run.

“We’ve run 7500 simulations since we started so that’s around 16 and a half million core hours of computing.

“Now to put that into context, if you did that on a pretty sophisticated four-core laptop it’ll take you 471 years to do what we’ve done in developing this car.

That’s 2200 computing hrs per simulation.

https://www.f1-forecast.com/pdf/F1-File ... 2e_all.pdf

It’s 14 years old now, but Honda does a fantastic job explaining how yaw, tire deformation, pitch, and internal flows effect all of this as well and you have to be able to model all that stuff.

[n_anirudh]

To the question of legality, it should pass, ML or DNN is just matrix operation and few more math operations and I dont see how FIA can police that.

If one looks up instant NGP from Nvidia, the renders are so life-like, and this is done using NN. This is similar to what Force India/Racing point did with some clever photogrammetry one can quickly re-create the 3D model of an F1 car or a part and evaluate how it works on their car. Even a copy may be a good starting point for an optimisation run

[dans79]

To the question of legality, it should pass, ML or DNN is just matrix operation and few more math operations and I dont see how FIA can police that.

Because their is bookkeeping and revision etc, it's in the regulations.