Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 


Forgot your password?
Close
typodupeerror
×
AI Google Technology

DeepMind Has Trained an AI To Control Nuclear Fusion (wired.com) 75

Posted by msmash from the closer-look dept.
The Google-backed firm taught a reinforcement learning algorithm to control the fiery plasma inside a tokamak nuclear fusion reactor. From a report: The inside of a tokamak -- the doughnut-shaped vessel designed to contain a nuclear fusion reaction -- presents a special kind of chaos. Hydrogen atoms are smashed together at unfathomably high temperatures, creating a whirling, roiling plasma that's hotter than the surface of the sun. Finding smart ways to control and confine that plasma will be key to unlocking the potential of nuclear fusion, which has been mooted as the clean energy source of the future for decades. At this point, the science underlying fusion seems sound, so what remains is an engineering challenge.

That's where DeepMind comes in. The artificial intelligence firm, backed by Google parent company Alphabet, has previously turned its hand to video games and protein folding, and has been working on a joint research project with the Swiss Plasma Center to develop an AI for controlling a nuclear fusion reaction. In stars, which are also powered by fusion, the sheer gravitational mass is enough to pull hydrogen atoms together and overcome their opposing charges. On Earth, scientists instead use powerful magnetic coils to confine the nuclear fusion reaction, nudging it into the desired position and shaping it like a potter manipulating clay on a wheel. The coils have to be carefully controlled to prevent the plasma from touching the sides of the vessel: this can damage the walls and slow down the fusion reaction. (There's little risk of an explosion as the fusion reaction cannot survive without magnetic confinement).

But every time researchers want to change the configuration of the plasma and try out different shapes that may yield more power or a cleaner plasma, it necessitates a huge amount of engineering and design work. Conventional systems are computer-controlled and based on models and careful simulations, but they are, Ambrogio Fasoli, director of the Swiss Plasma Center at Ecole Polytechnique Federale de Lausanne in Switzerland. says, "complex and not always necessarily optimized." DeepMind has developed an AI that can control the plasma autonomously. A paper published in the journal Nature describes how researchers from the two groups taught a deep reinforcement learning system to control the 19 magnetic coils inside TCV, the variable-configuration tokamak at the Swiss Plasma Center, which is used to carry out research that will inform the design of bigger fusion reactors in the future.
This discussion has been archived. No new comments can be posted.

DeepMind Has Trained an AI To Control Nuclear Fusion More

DeepMind Has Trained an AI To Control Nuclear Fusion

Comments Filter:

  • Let me be the first to say (Score:4, Funny)

    by Tablizer ( 95088 ) on Wednesday February 16, 2022 @01:10PM (#62273477) Journal

    Just Noooo!

    • Excellent application of heuristics. Just lovely

      • Re: (Score:1)

        by Tablizer ( 95088 )

        AI may be useful for pointing out if something looks out of whack (beyond normal checks), but controlling operations itself is too risky, at least until the technology becomes time and road tested as a critical component.

        • If you make it to the end of the summary, the AI is merely helping set up test scenarios. Once they find a design that works, they'll be able to design a static system that just works. Classical physics ought to be enough to finish the design, so it would be a fairly predictable reaction.

          • Re: (Score:3, Informative)

            by Tablizer ( 95088 )

            In that case, the headline is misleading: "DeepMind Has Trained an AI To Control Nuclear Fusion"

            Perhaps it should be, "DeepMind AI Helps Design Nuclear Fusion Controller".

            Using AI to help design a car engine is very different than using AI to drive the car itself. But few expect headlines to be accurate these days, because click-bait sells ads, and the scary version does that better.

            • It would be controlling the magnets on the prototype, so the headline would be correct. They just probably won't need the AI when they go to manufacturing.

              • Re: (Score:1)

                by Tablizer ( 95088 )

                Most readers would interpret the headline as being intended for "live" control. It's still possible to be technically accurate but misleading.

        • Nothing is too risky if there is a killing to be made. It's the human condition.

        • You just need to define the parameters of the controls so even if Deep Mind goes crazy and produces the worst possible outputs, then it won't cause a disaster. Inside those parameters, it's fine to let an AI control things.

          • Re: (Score:2)

            by Tablizer ( 95088 )

            > You just need to define the parameters of the controls so even if Deep Mind goes crazy and produces the worst possible outputs, then it won't cause a disaster. Inside those parameters, it's fine to let an AI control things.

            For complex systems there's often ways to bleep things up that can't be anticipated. Maybe some odd feedback loop can create unexpected vibrations or bubbles in the pipes or who knows what.

            And don't forget Stuxnet: it wore out parts without being detected for a long time using creati

            • Yeah but "self destruction" isn't a serious problem with fusion.

              • Re: (Score:1)

                by Tablizer ( 95088 )

                By some accounts they can spew radiation far and wide under certain malfunctions. The first "batch" will probably have unexpected bugs; it is rocket science.

                • The first "batch" will probably have unexpected bugs; it is rocket science.

                  True point. Strictly speaking it's much, much harder than rocket science.

    • Or, as Randall Munroe writes about astatine in "What If?"...

      "There’s no material safety data sheet... If there were, it would just be the word “NO” scrawled over and over in charred blood".

    • Re: (Score:2)

      by shanen ( 462549 )

      You're just in a race to get the first joke in, eh?

      But the joke I was looking for would have been something around "What could possibly go wrong?", but at a higher level. That one is obviously hanging too low. And the funniest jokes are the ones you weren't looking for. (The recent story about Metamates had a number of good ones.)

  • Gee. What could could go wrong?

    On the upside, millions might end up suffering, but I got a good chuckle out of the idea.

    • You should at least make a little effort to answer your own rhetorical question before saying something silly.
    • Nothing, technically. Fusion plants don't work yet, so the worst thing that can happen with this is that they continue not working.
  • Next upgrade will have four robotic tentacles and a direct spinal interface.

    • Re: (Score:1)

      by Tablizer ( 95088 )

      Dude, put down your jporn

  • with a form of fusion, and the robots had all the energy they needed.

  • Wheres the line (Score:5, Insightful)

    by jacks smirking reven ( 909048 ) on Wednesday February 16, 2022 @01:18PM (#62273501)

    I mean, this makes sense, we've been using computers to control processes that operate faster than human reflexes can handle for, well, almost as long as computers have been around.

    I know I am in "old man yells at cloud" territory but at what point is this an "AI" and not just a really, really advanced PID controller?

    • Once they're done training the model it'll look like a fuzzy logic controller from the 80's.

    • Re:Wheres the line (Score:4, Insightful)

      by ljw1004 ( 764174 ) on Wednesday February 16, 2022 @02:06PM (#62273725)

      I know I am in "old man yells at cloud" territory but at what point is this an "AI" and not just a really, really advanced PID controller?

      AI is the standard industry term for "machine learning" - millions of weighted variables and deep reinforcement learning. Get over it.

      • at it's root, AI/machine learning is just complex non-linear correlation. in other word, a best fit model.
    • Yes, it's just an alternate method of control. It might be dumb "AI"(like all of "AI"), but it's solving a problem humans can't, which is providing effective function approximation in situations where the function is not known and generally not very "discoverable" using other methods.

    • ...at what point is this an "AI" and not just a really, really advanced PID controller?

      If you define a cost function and use a search algorithm to tune the gains, by most definitions a PID controller is AI.

      The flip side is, we can use the tools of "AI" to make some really nice controllers.

    • Re: (Score:2)

      by necro81 ( 917438 )

      I know I am in "old man yells at cloud" territory but at what point is this an "AI" and not just a really, really advanced PID controller?

      To a large extent, a PID controller is deterministic. That is, with a bit of math you can show what the frequency response of the system will be. You can spot where you may encounter resonances, and avoid them. You can demonstrate what your stability margin is.

      With most of what they call AI these days (neural nets of one sort or another), it's a black box. You

    • but at what point is this an "AI"

      It is an AI when it stars in a Gibson novel; until then you're just an idiot.

    • Humans cannot process information as fast as a computer can. Some computer control systems can have latencies under 1ms for making decisions (that means from input to controller to actuator). A human? F1 drivers (which have some of the fastest reaction times have 200ms to respond to decisions.

      • That's what we need in charge of a fusion reactor - the wrong decision made really, really fast.

        • Or with modern multicore and distributed systems, the decision arrived at via committee consensus.

      • A human? F1 drivers (which have some of the fastest reaction times have 200ms to respond to decisions.
        It is more 100ms, and below.

    • Unless they say "Strong AI" or something similar, it's just a really advanced PID controller.

    • Roughly when were the first digital computers + sensors capable of controlling physical processes industrially available? It's so infrastructural now, I bet the general public doesn't realize that "faster than human reflexes" has been around for much of their lives.

      Most people don't realize that what you see in a movie is the acting computers getting instructions from the director to keep it slow so the humans can see what's going on. But when it comes to coffee, you better make sure the AI is already e

    • I know I am in "old man yells at cloud" territory but at what point is this an "AI" and not just a really, really advanced PID controller?

      Well it is an advanced controller by definition. What makes you think it's a PID controller. That's a precise type of model with various limitations.

    • "but at what point is this an "AI" and not just a really, really advanced PID controller?"

      Indeed, what this does is to use many many simulation data points to approximate a value function, so in the end it is really a nonlinear optimal controller synthesized in a different way.

    • at what point is this an "AI" and not just a really, really advanced PID controller?

      At the point where they trained a neural network to accomplish the goal.

      Could you implement a PID controller using a neural network? Yes. Does that mean it's merely a PID controller? No because it's not based on the same concept even though it has the same result.

    • If Google, Apple, Facebook or Microsoft really gave a $hit about the climate they'd invest a few tens of billions in their own moonshot program to develop a working fusion reactor in a timeframe sooner than the "10-20 years away" we've had for the past 70 years. If fusion can really ever truly work, that would be by far the most effective way they could spend that chump change they have sitting in offshore bank accounts collecting dust.

      Or maybe this is yet another industry too incompetent to support itself

  • "Keep calm! It's only a spike! It'll soon stabilize!"

  • A well trained AI takes 'a pretty decent' to a 'good' choice.
    Never a perfect choice.
    To control such facilities, fully deterministic software is needed:

    GOOD: Number in -> algorithm with proper physics equations -> decision out.

    Not:

    BAD: Number in -> something that seems to do a good job 99% of the time, because it scanned a polynomial space and found a decent local optimum -> decision out.

  • Why is fusion so hard! (Score:3)

    by Whateverthisis ( 7004192 ) on Wednesday February 16, 2022 @01:32PM (#62273553)
    Sheesh you humans. It's simple. Just change the gravitational constant of the universe thereby altering the mass of the fusion reaction.

    -Q.

  • Do you want a fusion powered Skynet? 'cause this is how you get fusion powered Skynet!

  • These computers with the ability to refine their algorithms is akin to self-sharpening knives. The knife can keep itself sharp and performs a task very well. However, it is not aware of the items it is cutting nor is it aware of the purpose for cutting the item. Cutting a carrot is no different than cutting a steak or a finger or a head of cow.

    The day it asks the question why, may be too late.

  • Is this another super battery article? All this innovation has me confused, yet I still have a lead acid battery in my gas powered car
    • You can swap it for a LiFePO4 battery and/or a bank of supercapacitors if you want to pay and know what you're doing. Lead-acid kinda sucks, but it's a cheap and mature technology.

  • its all similar to the way that Doc Ock controlled nuclear fusion. Let's just hope deepmind doesn't escape and start attacking the city.

  • Thanks Wired, but ... (Score:3)

    by fahrbot-bot ( 874524 ) on Wednesday February 16, 2022 @01:48PM (#62273625)

    Hydrogen atoms are smashed together at unfathomably high temperatures, ...

    This is science, we can fathom that.

  • The cooling on one of the CPU's fails, the CPU smokes, the AI glitches out and BOOM! Nuclear fusion to nuclear explosion!
  • Right, but what are they going to do with neutrons? Hydrogen-hydrogen fusion produces neutrons, and because they lack charge, they cannot be confined by an electromagnetic field. In current design, they hammer the surrounding structure, which is fine for an experimental setup, but less convincing for an industrial one.

    • Re: (Score:1)

      by 0dugo0 ( 735093 )

      On an industrial scale you recycle the damaged walls if you don't want industrial scale radioactive waste.

Slashdot Top Deals

Don't tell me how hard you work. Tell me how much you get done. -- James J. Ling

Close