Machine learning and AI for design

Here are our CFD links and discussions about aerodynamics, suspension, driver safety and tyres. Please stick to F1 on this forum.
Powy
19
Joined: Tue May 26, 2015 12:09 pm

Machine learning and AI for design

Post by Powy » Fri Sep 01, 2017 3:08 am

I've recently watched this fascinating talk on how Machine Learning and AI can help us with designs. For example, in the video you can see that they fed data to a machine learning algorithm and it came up with a frame for a car.

Does anyone know if teams are already using this in F1 ?



I can imagine it being useful for a compact packaging and chassis design to save weight (e.g. thinking back of the Sauber C32 in 2013 with the extremely small sidepods).

Greg Locock
138
Joined: Fri Jun 29, 2012 11:48 pm

Re: Machine learning and AI for design

Post by Greg Locock » Tue Sep 05, 2017 8:06 am

I'd hope that F1 uses this rather ancient technology. In 1982 the rear parcel shelf on some Audi model was optimised in fea. I didn't watch the ted life is too short. Commercial products such as optistruct have been around forever, and I wrote my own optimiser for spaceframes 15 years ago.

wgknestrick
1
Joined: Thu Sep 18, 2014 5:08 pm

Re: Machine learning and AI for design

Post by wgknestrick » Mon Oct 23, 2017 3:43 am

Video didn't work, but I can speculate what (iterative) tech you are referencing...

It only gets you so far (mainly for homogeneous materials), and IMO, the teams are well past that and onto algorithm based lattice structures via 3D printing. 3D printing is absolutely the future of F1 with respect to optimizing designs and creativity with microstructures. It basically removes all limits of manufacturing except general constraint of SLS process is about a .028-.030in min feature size, dia strut or lattice. You can print functional heatpipes, mechanisms, heat exchanger cores, mass dampers, oil passages, filters, baffles, etc right inside a solid exterior part......and no one would ever know it's in there (hint hint). Not to mention how easy it is to design incredibly strong joining structures to CF or overmold the 3d printed parts. If F1 is allowed per regs, they are probably 3D printing a metal part.

I work with this kind of tech in the medical implant field and you cannot believe how strong (and light) these structures are in Ti SLS material.
Image

Some sort of HE core...maybe an F1 intercooler
Image

This software (or their own custom versions) is where most teams are probably working now. Machine learning certainly has it's place in F1 though. Strategies, fuel/ harvesting mapping, etc. With F1, if it exists and can give them ANY advantage, just assume they are using it.
https://www.autodesk.com/products/withi ... l/overview

The complexity of the lattices quickly outpace the limitations of modern CAD software and you need to separately mesh your CAD designs in a post processing step to ready the STL files for printing.

ChrisDanger
41
Joined: Wed Mar 30, 2011 8:59 am

Re: Machine learning and AI for design

Post by ChrisDanger » Tue Oct 24, 2017 7:33 am

This is actually very interesting. It's a pity it doesn't seem to be getting much attention. The novelty here is that AI, in the form of some kind of genetic algorithm, performs the chassis design purely from inputs of boundary conditions in the form of forces at certain points like suspension and engine mounts. The result is quite different to what a human could conceive, is quite lightweight, and is certainly only able to be manufactured by 3D printing technology.

This is the relevant part: https://youtu.be/aR5N2Jl8k14?t=759

godlameroso
155
User avatar
Joined: Sat Jan 16, 2010 8:27 pm
Location: Miami FL

Re: Machine learning and AI for design

Post by godlameroso » Tue Oct 24, 2017 6:39 pm

It also has the nice side effect of creating generally more organic and streamlined shapes, as nature tends to do, which is pleasing to the eye.
The height of cultivation is really nothing special. It is merely simplicity; the ability to express the utmost with the minimum. Mr.Lee

Greg Locock
138
Joined: Fri Jun 29, 2012 11:48 pm

Re: Machine learning and AI for design

Post by Greg Locock » Wed Oct 25, 2017 12:10 am

I like the chassis frame in that TED.

For those of you who are students or pirates, grab a copy of Altair Hyperworks and do the Optistruct tutorial on a suspension arm.

For those of you who aren't pirates, I once designed a subframe by assembling a set of forces at each mounting point, and connected them all with equal size tubes. I then ran the forces, and knocked out all the tubes that were blue in the stress plots. Then I did it again. Then I beefed up the red ones. And so on and so forth. It was at once apparent that a belly pan was the single most efficient structural component, sadly the sump and diff were in the way.

godlameroso
155
User avatar
Joined: Sat Jan 16, 2010 8:27 pm
Location: Miami FL

Re: Machine learning and AI for design

Post by godlameroso » Fri Oct 27, 2017 10:27 pm

As far as aerodynamics are concerned, does anyone think these algorithms would come up with more and more bird feather type designs?

I look at the current crop of cars and see all the small slots and winglets and think to myself this car is starting to look more and more like a bird. All the little attachments looking like feathers, and if you think about it, it makes a lot of sense. If you want to control the wind you should copy the experts who have been evolving the ability as long as humans have been around on this planet(birds).

Tell me you can't see the resemblance between the slots and the edge of this bird wing.

Image

Image
The height of cultivation is really nothing special. It is merely simplicity; the ability to express the utmost with the minimum. Mr.Lee

roon
109
Joined: Sat Dec 17, 2016 6:04 pm

Re: Machine learning and AI for design

Post by roon » Fri Oct 27, 2017 10:55 pm

Perhaps, but I'm more inclined to compare an F1 car at speed to a fish in water. Could explore anatomy from that arena. I think the Reynolds numbers might be more comparable. Tuna have vortex generators along the edges of their tail, for example.

Image

Image

Manta rays have a very "motorsport" look to them.

Image

Edax
37
Joined: Tue Apr 08, 2014 9:47 pm

Re: Machine learning and AI for design

Post by Edax » Sat Oct 28, 2017 12:33 am

wgknestrick wrote:
Mon Oct 23, 2017 3:43 am
Video didn't work, but I can speculate what (iterative) tech you are referencing...

It only gets you so far (mainly for homogeneous materials), and IMO, the teams are well past that and onto algorithm based lattice structures via 3D printing. 3D printing is absolutely the future of F1 with respect to optimizing designs and creativity with microstructures. It basically removes all limits of manufacturing except general constraint of SLS process is about a .028-.030in min feature size, dia strut or lattice. You can print functional heatpipes, mechanisms, heat exchanger cores, mass dampers, oil passages, filters, baffles, etc right inside a solid exterior part......and no one would ever know it's in there (hint hint). Not to mention how easy it is to design incredibly strong joining structures to CF or overmold the 3d printed parts. If F1 is allowed per regs, they are probably 3D printing a metal part.

I work with this kind of tech in the medical implant field and you cannot believe how strong (and light) these structures are in Ti SLS material.
https://res.cloudinary.com/engineering- ... qs3xna.jpg

Some sort of HE core...maybe an F1 intercooler
http://www.eurekamagazine.co.uk/article ... _popup.jpg

This software (or their own custom versions) is where most teams are probably working now. Machine learning certainly has it's place in F1 though. Strategies, fuel/ harvesting mapping, etc. With F1, if it exists and can give them ANY advantage, just assume they are using it.
https://www.autodesk.com/products/withi ... l/overview

The complexity of the lattices quickly outpace the limitations of modern CAD software and you need to separately mesh your CAD designs in a post processing step to ready the STL files for printing.
I love additive manufacturing, but having worked with it for many years I am less bullish.

I can see where you're coming from but outside biotech additive manufacturing and prototyping it still has to live up its promise. And even in implants, while great progress is made in the academic world, the FDA is still trying to figure out how to ever certify this stuff, right?

The shape freedom for SLS is fantastic, but you pay a quite high price when it comes to engineering properties. Because you start with a powder you inevitabely end up with a high oxide content, some porosity, uncontrolled carbon content etc. Also you cannot texture the material like with forging. So you end up with a equalaxially grained material which performs poorly as compared to the traditional alloys ( strength, brittleness, thermal conductivity etc).

If you are used to working in low grade materials then SLS like the TiAlV you show might be OK. But for high end applications we are normally working with very expensive custom alloys. And then the performance gap of these printed materials is often just too big.

A second consideration is the quality (assurance) when you mill a component from a certified billet quality is quite easy to achieve. When you make a complex freeform structure from powder or rod you’ll often end up spending a lot of time looking at Xray tomographs or ultrasonic cross sections to make sure it came out OK.

You do see some printed parts in F1 though. McLaren quite often has printed ( SLS chopped carbon-carbon) test components on the car in FP ( but replaces them with traditional laminates for the final part and race)

As for topology optimization (I suspect the video is about). The sales pitch is great. Have the computer optimize a structure. The only thing is you have to set the boundary conditions for the optimization, based on the material characteristics, manufacturability etc. That takes a lot of (human) knowledge.

In the end it is more like a creative discussion between the computer and the engineer. The computer suggests a structure, the engineer will modify it based on his knowledge and they will continue in cycles until both are satisfied. I find it fascinating, and it is usefull, but it is still a long way from the computer designing a component. And that might be for the better ( otherwise we would have to fire a lot of guys).

flynfrog
Moderator
User avatar
Joined: Thu Mar 23, 2006 9:31 pm

Re: Machine learning and AI for design

Post by flynfrog » Sat Oct 28, 2017 6:05 pm

If you would like to play with machine learning for car design/ waste some time at work try Box car

http://boxcar2d.com/index.html

J.A.W.
52
Joined: Mon Sep 01, 2014 4:10 am
Location: Altair IV.

Re: Machine learning and AI for design

Post by J.A.W. » Sat Oct 28, 2017 10:09 pm

AI convergence on organic shapes - per flow dynamics?

Image

Image
Sturmbannfuehrer Dr von Braun sez..
"Oberste Prioritat hat es Londoner Terror zu vergelten. Und danach, der Mondflug!"

mep
34
User avatar
Joined: Sat Oct 11, 2003 2:48 pm
Location: Germany

Re: Machine learning and AI for design

Post by mep » Sun Oct 29, 2017 2:33 pm

Those feathers like wing elements seen on the cars are a result of many iterations of human design work and testing. So far, it does not really make sense to let a computer do it. The CAD systems used to generate such surfaces will require some human input. You can setup algorithms to change certain parameters but it will not design something like this from scratch. A surface designer will have generated such a surface in ca. 2 days.

3D printed parts are very common in F1. Mainly for wind tunnel testing but also race car parts are done by this. Some can even be spotted if you look carefully.

Zynerji
10
Joined: Wed Jan 27, 2016 3:14 pm

Re: Machine learning and AI for design

Post by Zynerji » Sun Oct 29, 2017 6:28 pm

mep wrote:
Sun Oct 29, 2017 2:33 pm
Those feathers like wing elements seen on the cars are a result of many iterations of human design work and testing. So far, it does not really make sense to let a computer do it. The CAD systems used to generate such surfaces will require some human input. You can setup algorithms to change certain parameters but it will not design something like this from scratch. A surface designer will have generated such a surface in ca. 2 days.

3D printed parts are very common in F1. Mainly for wind tunnel testing but also race car parts are done by this. Some can even be spotted if you look carefully.
Force India steering wheel comes to mind...

roon
109
Joined: Sat Dec 17, 2016 6:04 pm

Re: Machine learning and AI for design

Post by roon » Fri Nov 17, 2017 12:19 am