Engineers Use AI To Wrangle Fusion Power For the Grid (princeton.edu) 69
An anonymous reader quotes a report from Princeton Engineering: In the blink of an eye, the unruly, superheated plasma that drives a fusion reaction can lose its stability and escape the strong magnetic fields confining it within the donut-shaped fusion reactor. These getaways frequently spell the end of the reaction, posing a core challenge to developing fusion as a non-polluting, virtually limitless energy source. But a Princeton-led team composed of engineers, physicists, and data scientists from the University and the Princeton Plasma Physics Laboratory (PPPL) have harnessed the power of artificial intelligence to predict -- and then avoid -- the formation of a specific plasma problem in real time.
In experiments at the DIII-D National Fusion Facility in San Diego, the researchers demonstrated their model, trained only on past experimental data, could forecast potential plasma instabilities known as tearing mode instabilities up to 300 milliseconds in advance. While that leaves no more than enough time for a slow blink in humans, it was plenty of time for the AI controller to change certain operating parameters to avoid what would have developed into a tear within the plasma's magnetic field lines, upsetting its equilibrium and opening the door for a reaction-ending escape.
"By learning from past experiments, rather than incorporating information from physics-based models, the AI could develop a final control policy that supported a stable, high-powered plasma regime in real time, at a real reactor," said research leader Egemen Kolemen, associate professor of mechanical and aerospace engineering and the Andlinger Center for Energy and the Environment, as well as staff research physicist at PPPL. The research opens the door for more dynamic control of a fusion reaction than current approaches, and it provides a foundation for using artificial intelligence to solve a broad range of plasma instabilities, which have long been obstacles to achieving a sustained fusion reaction. The team published their findings in the journal Nature.
In experiments at the DIII-D National Fusion Facility in San Diego, the researchers demonstrated their model, trained only on past experimental data, could forecast potential plasma instabilities known as tearing mode instabilities up to 300 milliseconds in advance. While that leaves no more than enough time for a slow blink in humans, it was plenty of time for the AI controller to change certain operating parameters to avoid what would have developed into a tear within the plasma's magnetic field lines, upsetting its equilibrium and opening the door for a reaction-ending escape.
"By learning from past experiments, rather than incorporating information from physics-based models, the AI could develop a final control policy that supported a stable, high-powered plasma regime in real time, at a real reactor," said research leader Egemen Kolemen, associate professor of mechanical and aerospace engineering and the Andlinger Center for Energy and the Environment, as well as staff research physicist at PPPL. The research opens the door for more dynamic control of a fusion reaction than current approaches, and it provides a foundation for using artificial intelligence to solve a broad range of plasma instabilities, which have long been obstacles to achieving a sustained fusion reaction. The team published their findings in the journal Nature.
Re: (Score:1, Informative)
Your momma's so black when she gets out of the car the oil light comes on
Possibly, but (Score:2, Insightful)
It seems a lot like "This is a highly chaotic, rapidly fluctuating physical system... We understand that, and can't solve the problem. However, we don't understand computer science so we're going to assume AI can figure it out for us".
Re: (Score:2, Insightful)
Is that before or after the AI hallucinates and blows a hole where the reactor once was?
Re: (Score:2)
It seems a lot like "This is a highly chaotic, rapidly fluctuating physical system... We understand that, and can't solve the problem. However, we don't understand computer science so we're going to assume AI can figure it out for us".
The D-III National Fusion Facility is owned and ran by General Atomics and is in one of their buildings. General Atomics has a long standing relationship with the UCSD Supercomputer Center and they are literally within Walking Distance of the D-III Facility. If any of the researchers needed computer help it would have been easily available. I know, I work for General Atomics and used to work for a different division a couple of hundred feet away from the facility.
It is ran [Re:Possibly, but] (Score:2)
It is ran? You are part of the problem
Well, "ran" in Japanese means "chaos or tumult." [wikipedia.org] Saying "The fusion facility is ran" is a pun!
Re: (Score:2)
Did they give you a Pip-Boy?
Re: Possibly, but (Score:2)
Why would âoeunderstanding computer scienceâ allow you to manually write an accurate program that forecasts chaos? The missing part of the puzzle here is absolutely not CS.
Re: (Score:2)
The missing part of the puzzle here is absolutely not CS.
Machine learning is most definitely CS.
Many apparently chaotic systems are actually pseudo-chaotic.
They appear chaotic to us, but a deep learning system can predict their behavior well enough to be useful.
Re: (Score:2)
Many apparently chaotic systems are actually pseudo-chaotic.
Chaotic does not mean random. You cannot forecast the precise state of a chaotic system, but there are still rules in chaos. AI is a nice tool to discover some rules. We just have to hope it will not hallucinate them.
Re: (Score:1)
Nothing in the so-called "AI" is "computer science", it is statistics and computational algorithms, that is, math. You can easily tell by taking a look at the names index in your numerical methods book, all those bastards were mathematicians.
The "computer science" part is writing up a reasonably good API (which scientists almost always fuck up) and getting the compiler switches right, which is asymptotically trivial even for a doctoral student.
Re: Possibly, but (Score:2)
Re: (Score:2)
Yeah, I've heard that in some places of the world there's no physics, mathematics, chemistry or biology anymore, it is all "science".
Like I said, you can look at the name index of the book your "experts" learn their buzzwords from and check for yourself.
It is really easy nowadays :)
Re: (Score:2)
I see you've followed your advice and have stopped even using a username.
Which makes it plainly obvious that the drivel above is based on your own life story that you try to ascribe to me.
Sad.
Re: (Score:2)
Actually, the part where you make the machine generate "better crap" faster can be CS.
Re: (Score:2)
Why can't I troll all the computer scientists here from time to time?
[$(1)$]
Re: (Score:2)
You can. But you have to live with answers that may expose the trolling. Sorry about that.
Re: (Score:2)
Why sorry? It is par for the course.
Re: (Score:2)
I mean you could say the same thing about physics, engineering, and pretty much any STEM discipline, that they are just math. But just as physics is 'just' math applied to studying physical processes, computer science is math applied to improving the computational ability and efficiency of computers, which AI development certainly is.
Re: (Score:2)
Writing an API is engineering, at best, not science.
Actual computer science IS math. Sometimes it's not even applied math.
Re: (Score:3)
Not exactly my field but I have some passing familiarity with it. This would not be GPT type AI, but use of neural nets in their universal function capacity. Seems like the right tool for the job, to tame the crazy dynamics of a plasma. The challenge of course is how to train it. For now, it seems they are taking a machine learning approach using only the primitive sensor data as input. It would probably be much better to use physical models of plasma as a base and let the networks improve the dynamics from
Re: (Score:2)
I am partially talking out my ass, since I have never tried this myself ...
Indeed. Nothing you've said here makes any sense, but I appreciate the honestly. That's the right attitude to have when learning any subject.
The math really isn't all that difficult. It just looks scary when you're not used to the syntax or using math to communicate ideas. If you're interested machine learning, course in statistics will help a lot. As for neural networks, I ran across this [neuralnetw...arning.com] a while back, which is a very digestible introduction, well-suited for beginners. (That means there's very little
Re: (Score:2)
The math really isn't all that difficult. It just looks scary when you're not used to the syntax or using math to communicate ideas
You betray a complete lack of familiarity with fluid dynamics.
Re: (Score:2)
Hey if it works then by all means. To my layman's mind it seems like a pretty self-contained problem with relatively few parameters that AI might actually be useful at. With the amount of money that's already being thrown at fusion, this thing may be worth a try. But I'll believe that it's useful when it's producing a stable reaction.
Re: (Score:2)
Yep, that is pretty much what I read there too: "We can use AI Magic to solve this!"
Re: (Score:2)
Re: (Score:2)
I can see how it might seem like that to soemone who doesn't understand plasma physics or computer science.
Re: (Score:2)
The physics problem is a turbulent plasma. The mathematics to model it are nuts and not friendly to real time operation. You don't want to be inverting monster matrices in real time, for example. There has got to be better way. Oh yeah, there is, neural nets used as universal functions with training to predict and control the dynamics.
Re: (Score:2)
What do you mean, "I don't want to"? It is immaterial if I want it or not.
I must do it - I invert monstrous matrices because they drive the neutron fission in a nuclear reactor.
I must do it realtime and ensure operational safety, otherwise y'all get famous (and stupid) fake quotes like "not bad, not terrible".
Before you throw deep learning at a "problem", you should reconsider the problem you think you're solving.
Re: (Score:2)
What do you mean, "I don't want to"? It is immaterial if I want it or not.
I must do it - I invert monstrous matrices because they drive the neutron fission in a nuclear reactor.
I must do it realtime and ensure operational safety, otherwise y'all get famous (and stupid) fake quotes like "not bad, not terrible".
There are lots of cases where we have an accurate model of the physics, but the math to apply the model to a particular problem is not tractable. In those cases there's often a different model which has been derived empirically that models the behavior without having to calculate the underlying physical behaviors.
For example, using quantum mechanics, we can just about model a single hydrogen atom, anything much more complicated and the equations become intractable. If we want to solve for more complicated
Re:Desperation (Score:4, Informative)
neural nets used as universal functions
First, they're not "used as" they "are". The universal approximation theorem is the theoretical foundation for neural networks, after all. It's also a lot less impressive than you think.
Most people, well, anyone who has run across the term outside a classroom, don't understand what it means when we say that a NN is a universal function. They take the term at face-value and assume that you can do anything with a neural network than you can do with a computer. That simply isn't true. The word "universal" is more than a little misleading.
There are real limits here. "Universal approximation" only applies to functions that are continuous and compact. The function space of any individual network is further bound by its size/shape and activation function. In terms of computational power, universal approximation only guarantees that we can design a feed-forward neural network that can approximate any lookup table to an arbitrary degree of precision. It's not magic.
There has got to be better way. Oh yeah, there is, neural nets
Don't be so sure. Neural networks are expensive, both to train and operate. They're also not known for their efficiency. There's been a ridiculous amount of work put into matrix inversion already, so you'd be foolish not to look at traditional algorithms before you throw a lot of money at a NN.
Re: (Score:2)
Using neural nets you do (large, sparse) matrix inversions in training, not in real time evaluation. Also, though you didn't mention it, plasma control lives in the time domain. You don't want to be doing FFTs in real time either. Approximation baby, for the win.
Re: (Score:2)
He was talking about using NNs for large matrix inversion. There has been some research done, believe it or not, though mostly with RNNs.
Also, though you didn't mention it, plasma control lives in the time domain. You don't want to be doing FFTs in real time either. Approximation baby, for the win.
I'm not a physicist, so I can't speak to that side of things. I can tell you that real-time FFT is an actual thing, though there are other options, depending on your application, like sliding DFT, that are significantly faster than FFT.
Re: (Score:2)
"20 years": how original of you! Not.
Re: (Score:2)
My goal here isn't originality, sorry.
Best Available Tool (Score:3)
You know you've got there, when you think you need to apply "deep learning solution" to a physics problem.
What are you talking about? We use deep learning in just about every field of physics to analyse data. Deep learning is an incredibly powerful analysis technique when it comes to searching for a signal hidden in a pile of background. Rather than sit down and manually try to come up with some selection criteria for your signal we now train ML algorithms to do this for us. We still use physics models to develop variables that will be useful for separating signal and background but then we let the ML algorith
Re: (Score:2)
Yes, I'm in the trenches, and I see it every day. "Deep learning" is very helpful when you don't have a signal in your mess, but you do need to write that article.
But keep the faith.
Re: (Score:2)
Try this: Replace the term "deep learning" with "statistics". You should start to feel better almost immediately.
Re: (Score:2)
Unpossible.
I don't know what you imagine when you say "statistics". There is a bunch of "statisticses" that appears in many points of the design of a physics experiment that do not look anything like "deep learning".
Without being exhaustive, you need a model with appropriate observables (statistics), estimators derived properly from these (statistics) that are shown to have desirable properties (more statistics), then an experiment design with appropriate characteristics to measure those (even more statisti
Re: (Score:2)
When I hear "deep learning" I know what it is
You really don't.
Re: (Score:2)
Sure, Jan.
Re: (Score:2)
Lol.
Here's the equation for a generalized linear model, of which most of the statistical techniques you've ever heard of are a special case:
Y = g^{-1}(XB)
Here is the equation for a layer of an artificial neural network:
Y = a(XW)
They're identical, except for which letters are typically chosen to write the equation. Deep learning means you take that thing and stack more than two of them.
Re: (Score:2)
Yes, I'm in the trenches...
I'm sure you are but physics data analysis has progressed a lot since the time of the first world war and you need to catch up!
Machine learning is an incredibly powerful analysis technique that helps us to get the most out of the data we have. We know this because ~30 years ago when I was a grad student ML was brand new in physics and people were very sceptical of it. So, at the time, we did analyses both ways: the "traditional way" with human-picked cuts and the ML-way with (at the time) neutral nets a
Re: Desperation (Score:2)
The only thing AI is doing to the grid (Score:2, Insightful)
is burning countless gigawatt-hours of energy to create fake product reviews on Amazon and deepfakes on Youtube.
Not counting the insane amounts of energy to power the servers that serve up the relentless onslaught of fucking AI news all day, every day.
Looks like buzzword bingo. (Score:1)
I remember watching a series of videos about fusion energy that had different experts in the field give an hour or so presentation on their work on fusion reactors. Those videos have to be close to 20 years old by now. In watching one of them I saw this expert on fusion explain some of the math involved. If I recall correctly the energy out depends on the volume of the plasma, and then energy in depend on the area. So if the plasma where the fusion occurs is spherical then energy out is dependent on the
Re: (Score:2)
Size: a hydrogen bomb is not very large, I would guess much less than ten feet. And I would guess that the fusion takes place in a much smaller part of that bomb. Of course with all the energy out--far more at a fraction of a second than you would want from a fusion reactor--the resulting explosion expands to take up all the room available...
Re: (Score:1)
A hydrogen bomb has a fission stage and a fusion stage, so not quite what people are looking for here. It seems the point of a fusion reactor is to avoid the need for fission.
I suppose we could construct a power plant that has a fission reactor to feed heat, pressure, and neutrons for a fusion reaction but that hardly seems efficient. Why not just skip the fusion part and use the heat and pressure for producing electricity, and the neutrons to breed more fission fuel?
Re: (Score:2)
Fusion releases energy, so if you could use the output of a fission reactor to efficiently produce fusion, it would be well worth doing.
Problem is, you can't. The intensity of the energy from a fission bomb is important for triggering fusion, and the conditions only exist for a tiny fraction of a second, so if you want a significant amount of fusion it has to also be intense.
The challenge of controlled fusion is to make it happen slowly enough that you can still use the reactor afterward.
Chaos...or Kaos (Score:4, Insightful)
First, a statistical approach to chaos (which is what instability is) actually makes sense. It's not like you can work with accurate numbers in some kind of formula, the data points would be too vast (as the article says). Instead, recognize and classify the instability by matching what you "see" with known patterns (which is sort of what ordinary AI vision tools do), and apply a correction based on the type and (presumably) orientation of the recognized pattern.
The only alternative would be to send Maxwell Smart in to CONTROL the KAOS.
Re: (Score:2)
It depends what the consequences of it failing are. If it fails safe, you just lose say a gigawatt or two from the grid, okay, we can handle that. Happens all the time anyway, nuclear reactors SCRAM on fault, turbines break, transmission lines go down.
As long as the fusion reactor can recover reasonably cheaply and quickly, then using inherently unreliable AI to improve the odds is fine.
Re: (Score:3)
Neural networks used for control can be made more reliable by imposing constraints on the solution, as well. There's been some work on neural network-based drone control that uses Lipschitz regularization to keep the drones from crashing, like Neural Lander [yisongyue.com].
Appears (Score:5, Interesting)
Appears to be an actual statistical feedback problem, something neural networks are good at. I'd slot this firmly in the "task that a simple, narrow AI is the right tool for".
This isn't a hype-AI.
Pushing out net positivity (Score:2)
Re: (Score:2)
Progress indeed! (Score:2)
So from now on, instead of saying, "fusion is always 30 years away", should we change it to, "fusion is always 25 years away"?
Using AI to control nuclear reaction... (Score:1)
... What could go wrong ?
Back in the old days... (Score:2)
Doc Ock (Score:2)
Re: (Score:2)
This.
Couldn't this have been done with a simple GA? (Score:2)
Is a neural net really needed for this. The instabilities are almost certainly periodic.
Now to reverse engineer what it spotted (Score:2)
If they just stop here, at letting AI do it for us, no one