Artificial Intelligence to Design Lightweight Parts

General engine tech -- Drag Racing to Circle Track

Moderator: Team

Kevin Johnson
HotPass
HotPass
Posts: 9365
Joined: Tue Nov 22, 2005 5:41 am
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by Kevin Johnson »

Rick! wrote: Sun Oct 13, 2019 7:31 pm
Kevin Johnson wrote: Sat Oct 12, 2019 1:18 pm Image
Just like in most cases of smart solutions that I have experienced, great ideas typically converge on very similar solutions.
Case in point, this optimized rocker arm. Using good fundamentals, the same answer can be reached through different means.
Except mine is better as it doesn't have stress attracting notches. :)
The solution space is not infinite in the case of this rocker arm. There will be a few local minimums and maximums in the solution surface with properly defined constraints, design variables and objective goal that funnel the solution into very similar load paths and geometry.
I don't need 50 or 100 generative solutions that are variations of a theme.

The one concerning issue of this kind of tech is having "operators" instead of good engineers applying AI.
If you don't know the fundamentals of physics and mechanics, how do you know the "answer" is right?
Image
Image
I think with varying starting dimensions that the solution surfaces will be different and so have differing maxima and minima.

I can imagine a CAD or drafting course for engineering students where the instructor presents an exemplar and asks them to emulate it, having a much more "funneled" solution. If two classes were shown different exemplars then the solution sets would be quite different.

It might be interesting to run a generated solution in reverse, i.e provide a larger target mass and have the program add mass. Repeat.

What would this tell you about the A.I.?
Driving Force Online: BREAKING NEWS—Ohio Governor Signs SEMA-Supported Vehicle Freedom Bill Into Law!
Prodarwin
New Member
New Member
Posts: 1
Joined: Tue Oct 15, 2019 8:19 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by Prodarwin »

I registered so I could see the pics in this thread. Now that I have I realize the question in the second post hasn't really been addressed.

Can someone explain to me what is 'AI' about the original images?

The images Rick! posted are from Inspire I believe (or another Altair product), which is an iterative solver, not artificial intelligence. Artificial Intelligence would be a substitute for human intelligence/decision making, not "simply" solving iterations and choosing the best performer for the given constraints over and over again.
SchmidtMotorWorks
Vendor
Posts: 11003
Joined: Mon Apr 11, 2005 2:30 am
Location: CA

Re: Artificial Intelligence to Design Lightweight Parts

Post by SchmidtMotorWorks »

What is shown can be accomplished with topology optimization.
AI falls into multiple descriptions. One distinction is the type of algorithms applied.
The difference between what is done with topology optimization and what can be done with a neural network is more of a grouping of style than capability. They are after all both optimization routines.
Where AI is going beyond topology optimization is that it can include a wide range of influences and merge them into the optimization cycle.
Parameters such as manufacturing method, material, quantity required, time schedules, tooling cost, supplier schedules...the list is only limited by your resources to define the parameters and optimization.
Some people think that AI is unique in that it can avoid entrapment in local extremes, but there are other algorithms than can do that too.
The line between AI and other optimization schemes is not sharp.
Helping to Deliver the Promise of Flying Cars
Kevin Johnson
HotPass
HotPass
Posts: 9365
Joined: Tue Nov 22, 2005 5:41 am
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by Kevin Johnson »

I am guessing that the "challenge" or "question" I am posing is unusual though I am not sure why.

If you have a program that optimizes one variable then theoretically it should be possible to reverse the precise steps followed and once again "arrive" at the initial starting point.

As the software is a sort of black box, to avoid cheating, one should only be cognizant of the initial input and output and then compare these over multiple forward and reverse operations.

A more common usage is in translation programs where you can translate a phrase into the target language from English and then reverse by translating the result back into English from the target language.

English: I want to go to the store; Target language Russian: я хочу пойти в магазин

Russian: я хочу пойти в магазин ; Target language English: I want to go to the store
Driving Force Online: BREAKING NEWS—Ohio Governor Signs SEMA-Supported Vehicle Freedom Bill Into Law!
SchmidtMotorWorks
Vendor
Posts: 11003
Joined: Mon Apr 11, 2005 2:30 am
Location: CA

Re: Artificial Intelligence to Design Lightweight Parts

Post by SchmidtMotorWorks »

Kevin Johnson wrote: Fri Oct 18, 2019 11:59 am I am guessing that the "challenge" or "question" I am posing is unusual though I am not sure why.

If you have a program that optimizes one variable then theoretically it should be possible to reverse the precise steps followed and once again "arrive" at the initial starting point.
Can you explain why you think that?

Two languages are not analogous to two shapes.
Helping to Deliver the Promise of Flying Cars
Kevin Johnson
HotPass
HotPass
Posts: 9365
Joined: Tue Nov 22, 2005 5:41 am
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by Kevin Johnson »

SchmidtMotorWorks wrote: Sat Oct 19, 2019 4:24 am
Kevin Johnson wrote: Fri Oct 18, 2019 11:59 am I am guessing that the "challenge" or "question" I am posing is unusual though I am not sure why.

If you have a program that optimizes one variable then theoretically it should be possible to reverse the precise steps followed and once again "arrive" at the initial starting point.
Can you explain why you think that?

Two languages are not analogous to two shapes.
The translation program processes information which is represented in each language. If the translation program is consistent, it will represent the information in the same way forwards and backwards into the target languages.

Just now I performed an experiment with Google translator:

I am going to a restaurant to eat an expensive and delicious meal

--> Я иду в ресторан, чтобы поесть дорогую и вкусную еду

Я иду в ресторан, чтобы поесть дорогую и вкусную еду

--> I go to a restaurant to eat expensive and delicious food
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

I am going to a restaurant to eat an expensive and delicious meal and then I will drive my car home.

--> Я иду в ресторан, чтобы поесть дорогую и вкусную еду, а потом отвезу машину домой

Я иду в ресторан, чтобы поесть дорогую и вкусную еду, а потом отвезу машину домой

--> I go to a restaurant to eat expensive and tasty food, and then I will drive the car home
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Once I had obtained a translation, Google was subsequently consistent in yielding the forward and reverse results which suggests the information is stored.

Clearly, Google is not consistent in selection of adjectives, cases and possessives, etc.

~~~~~~~~~~~~~~~~~~~~~~`

Perhaps the problem would be more clearly stated in Fourier Analysis:

When a signal is analyzed and a transform developed, is the inverse of the transform unique to the original signal? I do not think so, generally.

~~~~~~~~~~~~~~~~~~~~~~~~~

When the surface of an object is optimized in some rule-based system, is it normally the case that the inverse application of the rule-based system will yield the original exemplar or will the original exemplar merely be one possible element in a set (infinite in cardinality?)? My feeling is that this is the genesis of a so-called inverse problem.
Driving Force Online: BREAKING NEWS—Ohio Governor Signs SEMA-Supported Vehicle Freedom Bill Into Law!
hoffman900
HotPass
HotPass
Posts: 3423
Joined: Sat Feb 23, 2013 5:42 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by hoffman900 »

Kevin,

You should probably listen to this Podcast: http://lineardigressions.com/episodes/2 ... ranslation

It’s a good podcast about machine learning. They play the naive host / expert, but they’re both very sharp. Most would be shocked here how far technology has progressed, especially considering some are still hung up on the carburetor.

The paper: https://arxiv.org/pdf/1609.08144.pdf

Also this: https://www.newyorker.com/magazine/2019 ... new-yorker

It’s a fantastic article.
-Bob
GARY C
HotPass
HotPass
Posts: 6301
Joined: Tue May 14, 2013 10:58 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by GARY C »

hoffman900 wrote: Sat Oct 19, 2019 1:19 pm Kevin,

You should probably listen to this Podcast: http://lineardigressions.com/episodes/2 ... ranslation

It’s a good podcast about machine learning. They play the naive host / expert, but they’re both very sharp. Most would be shocked here how far technology has progressed, especially considering some are still hung up on the carburetor.

The paper: https://arxiv.org/pdf/1609.08144.pdf

Also this: https://www.newyorker.com/magazine/2019 ... new-yorker

It’s a fantastic article.
Probably because they are still out performing EFI at a fraction of the cost for both equipment and tune.
Please Note!
THE ABOVE POST IN NO WAY REFLECTS THE VIEWS OF SPEED TALK OR IT'S MEMBERS AND SHOULD BE VIEWED AS ENTERTAINMENT ONLY...Thanks, The Management!
User avatar
BigBlockMopar
Momentary Specialist
Momentary Specialist
Posts: 336
Joined: Fri Mar 25, 2016 8:53 pm
Location: Netherlands
Contact:

Re: Artificial Intelligence to Design Lightweight Parts

Post by BigBlockMopar »

I would like to see the result of this on something more complex like a crank, or something else that rotates/accelerates.
ClassAct
Guru
Guru
Posts: 1018
Joined: Wed Dec 05, 2018 11:55 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by ClassAct »

GARY C wrote: Sat Oct 19, 2019 3:02 pm
hoffman900 wrote: Sat Oct 19, 2019 1:19 pm Kevin,

You should probably listen to this Podcast: http://lineardigressions.com/episodes/2 ... ranslation

It’s a good podcast about machine learning. They play the naive host / expert, but they’re both very sharp. Most would be shocked here how far technology has progressed, especially considering some are still hung up on the carburetor.

The paper: https://arxiv.org/pdf/1609.08144.pdf

Also this: https://www.newyorker.com/magazine/2019 ... new-yorker

It’s a fantastic article.
Probably because they are still out performing EFI at a fraction of the cost for both equipment and tune.

LOL. There is always that!!
hoffman900
HotPass
HotPass
Posts: 3423
Joined: Sat Feb 23, 2013 5:42 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by hoffman900 »

GARY C wrote: Sat Oct 19, 2019 3:02 pm
hoffman900 wrote: Sat Oct 19, 2019 1:19 pm Kevin,

You should probably listen to this Podcast: http://lineardigressions.com/episodes/2 ... ranslation

It’s a good podcast about machine learning. They play the naive host / expert, but they’re both very sharp. Most would be shocked here how far technology has progressed, especially considering some are still hung up on the carburetor.

The paper: https://arxiv.org/pdf/1609.08144.pdf

Also this: https://www.newyorker.com/magazine/2019 ... new-yorker

It’s a fantastic article.
Probably because they are still out performing EFI at a fraction of the cost for both equipment and tune.
Lol. On a static dyno, with no gforces, and on engines that are set up for carburetors and just have a compromised EFI system pasted on?

I know everyone bemoans the lack open rule sets, and unshockingly, where they exist, carburetors haven’t been a thing in a long time. Furthermore, like Pro Stock and NASCAR, they are limited by rules on what they can do.

This has been hashed over and over again. The F1 teams measure wave speed, which is a function of density, which is effected by atomization, and can work the injector design to maximize density. Try that with a carburetor. Watch Kaase’s clear manifold video to see why that doesn’t work, and that’s without gforces acting on it!

One thing I like about Speedtalk is discussing performance and going faster, not the same old vintage recipes doing the same old power numbers. This place use to be like F1technical at times. Less and less like that as time goes on.
-Bob
hoffman900
HotPass
HotPass
Posts: 3423
Joined: Sat Feb 23, 2013 5:42 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by hoffman900 »

BigBlockMopar wrote: Sat Oct 19, 2019 4:30 pm I would like to see the result of this on something more complex like a crank, or something else that rotates/accelerates.
Not exactly there yet, but it's interesting to see the evolution of the F1 rod / piston (which the NA vs. Turbo has some influence on).

https://www.f1technical.net/forum/viewt ... =4&t=27372
-Bob
GARY C
HotPass
HotPass
Posts: 6301
Joined: Tue May 14, 2013 10:58 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by GARY C »

hoffman900 wrote: Sat Oct 19, 2019 6:28 pm
GARY C wrote: Sat Oct 19, 2019 3:02 pm
hoffman900 wrote: Sat Oct 19, 2019 1:19 pm Kevin,

You should probably listen to this Podcast: http://lineardigressions.com/episodes/2 ... ranslation

It’s a good podcast about machine learning. They play the naive host / expert, but they’re both very sharp. Most would be shocked here how far technology has progressed, especially considering some are still hung up on the carburetor.

The paper: https://arxiv.org/pdf/1609.08144.pdf

Also this: https://www.newyorker.com/magazine/2019 ... new-yorker

It’s a fantastic article.
Probably because they are still out performing EFI at a fraction of the cost for both equipment and tune.
Lol. On a static dyno, with no gforces, and on engines that are set up for carburetors and just have a compromised EFI system pasted on?

I know everyone bemoans the lack open rule sets, and unshockingly, where they exist, carburetors haven’t been a thing in a long time. Furthermore, like Pro Stock and NASCAR, they are limited by rules on what they can do.

This has been hashed over and over again. The F1 teams measure wave speed, which is a function of density, which is effected by atomization, and can work the injector design to maximize density. Try that with a carburetor. Watch Kaase’s clear manifold video to see why that doesn’t work, and that’s without gforces acting on it!

One thing I like about Speedtalk is discussing performance and going faster, not the same old vintage recipes doing the same old power numbers. This place use to be like F1technical at times. Less and less like that as time goes on.
Ever wonder why they mandated EFI as opposed to "allowing" EFI vs Carb?
Please Note!
THE ABOVE POST IN NO WAY REFLECTS THE VIEWS OF SPEED TALK OR IT'S MEMBERS AND SHOULD BE VIEWED AS ENTERTAINMENT ONLY...Thanks, The Management!
hoffman900
HotPass
HotPass
Posts: 3423
Joined: Sat Feb 23, 2013 5:42 pm
Location:

Re: Artificial Intelligence to Design Lightweight Parts

Post by hoffman900 »

GARY C wrote: Sat Oct 19, 2019 6:49 pm
hoffman900 wrote: Sat Oct 19, 2019 6:28 pm
GARY C wrote: Sat Oct 19, 2019 3:02 pm
Probably because they are still out performing EFI at a fraction of the cost for both equipment and tune.
Lol. On a static dyno, with no gforces, and on engines that are set up for carburetors and just have a compromised EFI system pasted on?

I know everyone bemoans the lack open rule sets, and unshockingly, where they exist, carburetors haven’t been a thing in a long time. Furthermore, like Pro Stock and NASCAR, they are limited by rules on what they can do.

This has been hashed over and over again. The F1 teams measure wave speed, which is a function of density, which is effected by atomization, and can work the injector design to maximize density. Try that with a carburetor. Watch Kaase’s clear manifold video to see why that doesn’t work, and that’s without gforces acting on it!

One thing I like about Speedtalk is discussing performance and going faster, not the same old vintage recipes doing the same old power numbers. This place use to be like F1technical at times. Less and less like that as time goes on.
Ever wonder why they mandated EFI as opposed to "allowing" EFI vs Carb?
Yeah, because they would have been all EFI. lol.

The rules are set in place to cheapen the costs of the change and to benefit one sponsor, Holley. Holley drove that entire rule set.

Otherwise they would all be high pressure systems, with possibly multiple injectors, and ECU's from Magneti Marelli, Mclaren, Bosch, etc.

The tunnel ram, dual carburetor manifold was the peak of development for the carburetor. What they have now isn't even close to the peak of development for EFI or touch upon its capabilities.

Try this with a carburetor: https://www.enginelabs.com/news/f1-tech ... -ignition/

What is the basis of your argument and why are you so amendment to defend it?
-Bob
oldjohnno
Pro
Pro
Posts: 395
Joined: Mon Feb 01, 2010 9:58 pm
Location:
Contact:

Re: Artificial Intelligence to Design Lightweight Parts

Post by oldjohnno »

I've seen images of other things that have been designed artificially - one of a bridge comes to mind - where practical considerations like the use of stock materials and ease of manufacture have been ignored. And the thing that strikes me with them is their resemblance to bones - when there are no other constraints other than bulk and strength-to-mass the results seem to appear very bone-like.
Perfectionism is the enemy of actually getting shit done.
Post Reply