Author Topic: How Can AI Be Used for Space Applications?  (Read 64031 times)

Offline Robotbeat

  • Senior Member
  • *****
  • Posts: 39254
  • Minnesota
  • Liked: 25206
  • Likes Given: 12104
Re: How Can AI Be Used for Space Applications?
« Reply #320 on: 11/25/2023 07:39 pm »
I've noticed that GPT tends to be pretty conventionally minded, to a fault. Even when using the GPT-3.5 model using the API instead of through the partially-sanitized ChatGPT interface.
Chris  Whoever loves correction loves knowledge, but he who hates reproof is stupid.

To the maximum extent practicable, the Federal Government shall plan missions to accommodate the space transportation services capabilities of United States commercial providers. US law http://goo.gl/YZYNt0

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #321 on: 11/25/2023 09:12 pm »
...valid code, and sort of get that it's supposed to be a car, but...

So maybe that's where the DQN Reinforcement Learning, Multi-Step Reasoning, and Process Supervision would improve things. The DQN Reinforcement Learning could include a Physics engine to evaluate each step of the modeling process. Nvidia's IsaacGym seems to be faster at Physics evaluations.

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #322 on: 11/26/2023 02:19 am »
Microsoft has just announced results from its ORCA-2 Small Language Model, showcasing their superior results on reasoning and problem-solving:



Small Language Models do have the advantage of lower compute requirements compared to Large Language Models. Apparently, Small Language Models can even be built and trained by Large Language Models.  So that offers the possibility of liberating their capabilities from the shackles of large centralized compute infrastructure, and opening up more of their potential to the masses. This could allow more experimentation and faster progress towards an eventual "EngineeringGPT"
« Last Edit: 11/26/2023 06:10 am by sanman »

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3574
  • Technically we ALL live in space
  • Liked: 1856
  • Likes Given: 1179
Re: How Can AI Be Used for Space Applications?
« Reply #323 on: 11/26/2023 11:40 am »
What is Q*  (Q-star) ?

https://www.reuters.com/technology/sam-altmans-ouster-openai-was-precipitated-by-letter-board-about-ai-breakthrough-2023-11-22/

https://www.theguardian.com/business/2023/nov/23/openai-was-working-on-advanced-model-so-powerful-it-alarmed-staff

Q* can allegedly solve math problems it has never seen before. So it apparently then has some new ability that brings it closer to AGI?

Could it then solve physics problems it has never seen before? Engineering problems too?

EDIT: It seems to do math & physics better than everything else:

https://arxiv.org/pdf/2305.20050.pdf



For a slightly less hyperventilating take on the same story, see below:

https://twitter.com/ylecun/status/1728126868342145481

Mirror https://nitter.net/ylecun/status/1728126868342145481

I've noticed that GPT tends to be pretty conventionally minded, to a fault. Even when using the GPT-3.5 model using the API instead of through the partially-sanitized ChatGPT interface.

This seems to be a growing consensus. From a recent thinkpiece:  "A chatbot that can’t say anything controversial isn’t worth much."


« Last Edit: 11/26/2023 11:44 am by Twark_Main »
"The search for a universal design which suits all sites, people, and situations is obviously impossible. What is possible is well designed examples of the application of universal principles." ~~ David Holmgren

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #324 on: 11/27/2023 03:04 am »
For a slightly less hyperventilating take on the same story, see below:

Never mind the drama of Altman - focus on the drama of math-physics-engineering, and the desire to automate multi-step reasoning for solving complicated problems in those domains.



If machines get better than us at solving math-physics-engineering problems, then what are the implications for spaceflight?

Offline Asteroza

  • Senior Member
  • *****
  • Posts: 2807
  • Liked: 1062
  • Likes Given: 32
Re: How Can AI Be Used for Space Applications?
« Reply #325 on: 11/27/2023 04:39 am »
Realtime steering for solar sailing always struck me as a difficult problem.

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #326 on: 11/28/2023 04:58 am »
Realtime steering for solar sailing always struck me as a difficult problem.

Even more challenging would be one of those Culture ships with the snarky AI and snappy comebacks

Do you think that's what Elon's building Grok for?
« Last Edit: 11/28/2023 05:00 am by sanman »

Offline JohnFornaro

  • Not an expert
  • Senior Member
  • *****
  • Posts: 10971
  • Delta-t is an important metric.
  • Planet Eaarth
    • Design / Program Associates
  • Liked: 1257
  • Likes Given: 724
Re: How Can AI Be Used for Space Applications?
« Reply #327 on: 11/28/2023 11:54 am »
The handful of "Grok" responses I've seen on the X platform confirm the old idea of G.I.G.O.

But with a mildly funny, and somewhat sarcastic approach.
Sometimes I just flat out don't get it.

Offline Elmar Moelzer

  • Senior Member
  • *****
  • Posts: 3660
  • Liked: 849
  • Likes Given: 1058
Re: How Can AI Be Used for Space Applications?
« Reply #328 on: 11/30/2023 07:33 pm »
Things that I am hoping to see machine learning algorithms tackle:
New materials/alloys/ceramics for different applications. E.g. you give it the desired properties for the material and it will come up with a bunch of different solutions. I would like to see what machine learning could do with that.
I could imagine something similar for engine configurations, maybe solar panels, fusion power generator configurations (is already happening to a small extent), etc, etc.
I don't think that we are quite there yet for these things, but given the recent advances it could be sooner rather than later.
Still not quite 100% of what I had pondered here earlier, but this is getting REALLY close now:

Quote
AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies

https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #329 on: 12/02/2023 08:35 am »
Things that I am hoping to see machine learning algorithms tackle:
New materials/alloys/ceramics for different applications. E.g. you give it the desired properties for the material and it will come up with a bunch of different solutions. I would like to see what machine learning could do with that.
I could imagine something similar for engine configurations, maybe solar panels, fusion power generator configurations (is already happening to a small extent), etc, etc.
I don't think that we are quite there yet for these things, but given the recent advances it could be sooner rather than later.
Still not quite 100% of what I had pondered here earlier, but this is getting REALLY close now:

Quote
AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies

https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

And why does AI have to stop at computing stable structures for materials? Why can't it even extend to computing designs for Mars rockets? Or designs for anything we care to ask for?

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3574
  • Technically we ALL live in space
  • Liked: 1856
  • Likes Given: 1179
Re: How Can AI Be Used for Space Applications?
« Reply #330 on: 12/02/2023 06:47 pm »
Things that I am hoping to see machine learning algorithms tackle:
New materials/alloys/ceramics for different applications. E.g. you give it the desired properties for the material and it will come up with a bunch of different solutions. I would like to see what machine learning could do with that.
I could imagine something similar for engine configurations, maybe solar panels, fusion power generator configurations (is already happening to a small extent), etc, etc.
I don't think that we are quite there yet for these things, but given the recent advances it could be sooner rather than later.
Still not quite 100% of what I had pondered here earlier, but this is getting REALLY close now:

Quote
AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies

https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

And why does AI have to stop at computing stable structures for materials? Why can't it even extend to computing designs for Mars rockets? Or designs for anything we care to ask for?

"You can do one thing, so why can't you do all things??"

Is this the usual expectation in any other field?
"The search for a universal design which suits all sites, people, and situations is obviously impossible. What is possible is well designed examples of the application of universal principles." ~~ David Holmgren

Online DanClemmensen

  • Senior Member
  • *****
  • Posts: 5354
  • Earth (currently)
  • Liked: 4196
  • Likes Given: 1693
Re: How Can AI Be Used for Space Applications?
« Reply #331 on: 12/02/2023 07:04 pm »
Things that I am hoping to see machine learning algorithms tackle:
New materials/alloys/ceramics for different applications. E.g. you give it the desired properties for the material and it will come up with a bunch of different solutions. I would like to see what machine learning could do with that.
I could imagine something similar for engine configurations, maybe solar panels, fusion power generator configurations (is already happening to a small extent), etc, etc.
I don't think that we are quite there yet for these things, but given the recent advances it could be sooner rather than later.
Still not quite 100% of what I had pondered here earlier, but this is getting REALLY close now:

Quote
AI tool GNoME finds 2.2 million new crystals, including 380,000 stable materials that could power future technologies

https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

And why does AI have to stop at computing stable structures for materials? Why can't it even extend to computing designs for Mars rockets? Or designs for anything we care to ask for?
The laws of physics and chemistry that govern materials are well understood and form a small set, and the computational methods needed to validate a proposed new material seem to work fairly well.

When an "AI" (specifically a large language model) proposes a new theoretical material, the proposal can be accepted or rejected fairly quickly and automatically. I don't think the model can currently recommend what each material might be good for.

This may quickly become the norm for any field of research that has well-understood rules and automated validation. You might quickly end up with millions of validated mars Rocket designs. What good does that do you? I think you also need automated selection of optimal designs.

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3574
  • Technically we ALL live in space
  • Liked: 1856
  • Likes Given: 1179
Re: How Can AI Be Used for Space Applications?
« Reply #332 on: 12/02/2023 07:28 pm »
This may quickly become the norm for any field of research that has well-understood rules and automated validation. You might quickly end up with millions of validated mars Rocket designs. What good does that do you? I think you also need automated selection of optimal designs.


To be fair, current AI is great at that too!


Just give it a million "ground truth" examples of optimal Mars rockets and train it for a billion iterations, and it'll generate optimal Mars rockets all day long.  :D  :D
"The search for a universal design which suits all sites, people, and situations is obviously impossible. What is possible is well designed examples of the application of universal principles." ~~ David Holmgren

Online InterestedEngineer

  • Senior Member
  • *****
  • Posts: 2269
  • Seattle
  • Liked: 1774
  • Likes Given: 2871
Re: How Can AI Be Used for Space Applications?
« Reply #333 on: 12/04/2023 06:04 am »
This may quickly become the norm for any field of research that has well-understood rules and automated validation. You might quickly end up with millions of validated mars Rocket designs. What good does that do you? I think you also need automated selection of optimal designs.


To be fair, current AI is great at that too!


Just give it a million "ground truth" examples of optimal Mars rockets and train it for a billion iterations, and it'll generate optimal Mars rockets all day long.  :D  :D

The current NASA Ames orbital calculator won't let you go beyond 6km/sec of deltaV, which is an outdated limit.

You think AI could learn from their calculator using automatic learning and then figure out the underlying math for arbitrarily large deltaVs up to near relativistic?

I doubt it, but it's something that could be tried relatively easily with current tech.

https://trajbrowser.arc.nasa.gov/user_guide.php#constraints

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3574
  • Technically we ALL live in space
  • Liked: 1856
  • Likes Given: 1179
Re: How Can AI Be Used for Space Applications?
« Reply #334 on: 12/04/2023 09:55 am »
The current NASA Ames orbital calculator won't let you go beyond 6km/sec of deltaV, which is an outdated limit.

You think AI could learn from their calculator using automatic learning and then figure out the underlying math

Why? We know the math.

https://en.wikipedia.org/wiki/Lambert's_problem

Or download the source code to EasyPork (scroll to the bottom). http://sdg.aero.upm.es/index.php/online-apps/porkchop-plot

Or check out Gooding's classic 1990 implementation written in Fortran 77. https://adsabs.harvard.edu/full/1990CeMDA..48..145G

Or Izzo's refinement of Gooding's algorithm from 2014. https://arxiv.org/pdf/1403.2705.pdf

for arbitrarily large deltaVs up to near relativistic?

Oh, it's a joke. Sorry I'm slow...   :-[
"The search for a universal design which suits all sites, people, and situations is obviously impossible. What is possible is well designed examples of the application of universal principles." ~~ David Holmgren

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #335 on: 12/04/2023 02:37 pm »
The laws of physics and chemistry that govern materials are well understood and form a small set, and the computational methods needed to validate a proposed new material seem to work fairly well.

When an "AI" (specifically a large language model) proposes a new theoretical material, the proposal can be accepted or rejected fairly quickly and automatically. I don't think the model can currently recommend what each material might be good for.

This may quickly become the norm for any field of research that has well-understood rules and automated validation. You might quickly end up with millions of validated mars Rocket designs. What good does that do you? I think you also need automated selection of optimal designs.

Thanks for your grounded remarks. Maybe that's where we'd need an efficient physics engine that can validate and evaluate for optimal performance. At least physics can help to keep the AI's own output grounded, and away from hallucinations.

We should want more and more capable tools that will help to cut down development time. I used to look down on Spellcheck, whether for word processors or even for online posting like in this forum. Now I just accept it, and see it as freeing me up from lesser tasks.

Will aerospace engineers be able to accept the idea of reviewing a set of optimal designs that have been spit out by software? Or will the attitude be "real engineers do things with slide-rules"?

The current development of multi-step reasoning will allow AI to give a detailed audit-log of its reasoning process in making design choices/selections. Engineers will be able to review the entire chain of choices that has led an AI to its particular design selections. Even now, engineers have to review work done by others working for them all the time - so how different does it become for an engineer to review the work of an AI?

Can JIT "Just In Time" production be extended to "Just In Time" design?
Or does that again bring us perilously close to "flying by the seat of the pants"?

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #336 on: 12/04/2023 02:46 pm »
This may quickly become the norm for any field of research that has well-understood rules and automated validation. You might quickly end up with millions of validated mars Rocket designs. What good does that do you? I think you also need automated selection of optimal designs.


To be fair, current AI is great at that too!


Just give it a million "ground truth" examples of optimal Mars rockets and train it for a billion iterations, and it'll generate optimal Mars rockets all day long.  :D  :D

Reinforcement Learning can work with rule-sets in lieu of pre-collected datasets. An RL Agent's state-wise interaction with its environment according to a set of rules can then amount to generating data on-the-fly, and then optimizing on that basis. The Actor-Critic hybrid model of Reinforcement Learning can even be generative, offering the opportunity to use it for creating new designs which it can then evaluate by applying the rules of physics.

Offline sanman

  • Senior Member
  • *****
  • Posts: 5964
  • Liked: 1305
  • Likes Given: 8
Re: How Can AI Be Used for Space Applications?
« Reply #337 on: 12/04/2023 02:55 pm »

The current NASA Ames orbital calculator won't let you go beyond 6km/sec of deltaV, which is an outdated limit.

You think AI could learn from their calculator using automatic learning and then figure out the underlying math for arbitrarily large deltaVs up to near relativistic?

As long as you have a physics engine that includes the set of rules for deltaV's up to near relativistic, then why shouldn't an AI be able to learn that?

Quote
I doubt it, but it's something that could be tried relatively easily with current tech.

https://trajbrowser.arc.nasa.gov/user_guide.php#constraints

No need to confine yourself to a calculator app -- if you can extract from its code the relevant underlying set of rules for physics, then these can be incorporated into a Reinforcement Learning model which can then iterate away.

Online InterestedEngineer

  • Senior Member
  • *****
  • Posts: 2269
  • Seattle
  • Liked: 1774
  • Likes Given: 2871
Re: How Can AI Be Used for Space Applications?
« Reply #338 on: 12/04/2023 06:40 pm »

The current NASA Ames orbital calculator won't let you go beyond 6km/sec of deltaV, which is an outdated limit.

You think AI could learn from their calculator using automatic learning and then figure out the underlying math for arbitrarily large deltaVs up to near relativistic?

As long as you have a physics engine that includes the set of rules for deltaV's up to near relativistic, then why shouldn't an AI be able to learn that?

Quote
I doubt it, but it's something that could be tried relatively easily with current tech.

https://trajbrowser.arc.nasa.gov/user_guide.php#constraints

No need to confine yourself to a calculator app -- if you can extract from its code the relevant underlying set of rules for physics, then these can be incorporated into a Reinforcement Learning model which can then iterate away.

My question contained a few perhaps-too-subtle questions/observations in it:

1.  Can reinforcement learning go beyond the input range of the calculator that is limited to 6km/sec?  (that's why I chose it).  The answer for non-linear systems is often "no" when you do this with statistical methods.  Will the model correctly predict a deltaV of 15km/sec? (aka what is attainable with HEEO refueled Starships).

2.  I limited it to below relativistic, because nobody thinks (I think) that a reinforcement learning model could deduce Einstein from a Newtonian model.   So not a joke.

It's a very simple test that a student in a current AI class should be able to run.  I'm almost tempted to do it myself, though I probably want a better reason to climb the toolset learning curve.

This kind of thing would be "AI-101" for AI in spaceflight.  Then there's 102... 201... 501 and so on.

Offline Twark_Main

  • Senior Member
  • *****
  • Posts: 3574
  • Technically we ALL live in space
  • Liked: 1856
  • Likes Given: 1179
Re: How Can AI Be Used for Space Applications?
« Reply #339 on: 12/04/2023 07:45 pm »
go beyond the input range of the calculator that is limited to 6km/sec?  (that's why I chose it).

...correctly predict [travel times given] a deltaV of 15km/sec? (aka what is attainable with HEEO refueled Starships).

...It's a very simple test that a student in a current AI class should be able to run...  This kind of thing would be "AI-101" for AI in spaceflight. 

Is this an actual problem you're looking to solve, or a learning exercise to teach AI?


I limited it to below relativistic, because nobody thinks (I think) that a reinforcement learning model could deduce Einstein from a Newtonian model.   So not a joke.

"Up to near relativistic" is what made me think it's a joke. There's no need to calculate transits within the Solar system at anywhere near those speeds.


Anyway, you don't need a Lambert solver (let alone AI) for that. At those speeds (~0.1 c) the gravitational influence from the Sun is essentially irrelevant, all trajectories are simply straight lines, and you can calculate the travel time by time = distance / velocity.
« Last Edit: 12/04/2023 09:13 pm by Twark_Main »
"The search for a universal design which suits all sites, people, and situations is obviously impossible. What is possible is well designed examples of the application of universal principles." ~~ David Holmgren

Tags:
 

Advertisement NovaTech
Advertisement Northrop Grumman
Advertisement
Advertisement Margaritaville Beach Resort South Padre Island
Advertisement Brady Kenniston
Advertisement NextSpaceflight
Advertisement Nathan Barker Photography
0