Comment by PicassoCTs | Hacker News Reader

Comment by PicassoCTs | Hacker News Reader

Parent Full thread

I have a question, i ve read a book recommended on here, about fusion, called the "The Future of Fusion Energy". (It was very good, can recommend for a general topic overview and history of progress).

But i do not understand the general approach to plasma-vessel construction, in particular, why it is so passive. Its either "bend" to orbital shape dicated by the plasma physics properties when constructing (Wendelstein, Stellerators etc.).

Or increase the containment vessels magnetic field strength (scaling up the reactor, high-temperature-superconductors) until it is capable to hold all the chaotic increasing fluctuation inside no matter what.

My Question: Is there not a more active, not brute force approach for containing the plasma? Sort of "Whack-The-mole" countering the escape-events as they happen in a Tokamak?

If i understood correctly, the system adds to itself constantly energy (its a reactor after all) and any chaotic butterfly wing flapping on the plasma current can shape up to a tornado breaching the containment and ending fusion.

I'm a PhD student in robotics at Carnegie Mellon working on exactly this. It's extremely challenging for a few reasons:

- the dataset is a mess. The experiments that have been conducted on the tokamak that we have access to were done for very many different reasons and under many different configurations of the machine so there is not a clear method for disambiguating what dynamical changes are due to differences in the system vs underlying dynamical truths

- the simulators available are very slow and not that accurate

- the physics is hard enough that it's not possible to develop a controller in closed form (obviously)

This implies that we need a version of reinforcement learning or model-predictive control that is substantially more robust and sample-efficient than currently exists. We're working on that but obviously it's an open research problem.

I'm assuming disruption mitigation is a big part of what you're looking at? Controlling a normal state plasma would be nice, but recognizing incipient disruptions is absolutely critical if ITER (or ARC) is to function at all. One unmitigated disruption could break ITER.

Yeah so the paper I linked on contextual bayesian optimization (https://papers.nips.cc/paper/2019/hash/7876acb66640bad41f1e1...) does a combination of controlling for beta and optimizing linear MHD stability --- which is part of the problem. One of our collaborators has been working precisely on disruption prediction and mitigation but it's on most of our minds: https://www.pppl.gov/news/2021/artificial-intelligence-helps....

Any papers you can link to on this? Sounds interesting.

will repost my comment from the other one:

Papers I would recommend from our collaboration on control of normalized plasma pressure: https://papers.nips.cc/paper/2019/hash/7876acb66640bad41f1e1...

plasma profile transport modeling: https://iopscience.iop.org/article/10.1088/1741-4326/abe08d/...

hybrid dynamical modeling of gross plasma quantities: https://arxiv.org/abs/2006.12682

uncertainty quantification for plasma dynamics: https://arxiv.org/abs/2011.09588

It's still early days for this work and for us but we're looking at pushing reinforcement learning in methods and engineering to solve this problem.

I thought the design of the Wendelstein was actually to be able to control the plasma more completely by leveraging modern computing. Its design allows for a lot more input than a simple Tokamak would. But I only have a very superficial understanding of this.

I don't understand why you're saying the Wendelstein follows the path of the plasma, it's not like plasma has a shape it wants to be in, you just sort of push it the way you want it to go, and wherever it goes you try and push it the other way again, to make it go in a circle.

What other shape than the Stellerator would someone choose if they wanted to use even more fine control over the plasma?

There is a fusion startup called CT Fusion that has an innovation in this vein: [warning: not a plasma physicist, this is just what I understand from talking to actual physics nerds who follow this] instead of using giant magnets for plasma confinement, they are using a continuously variable magnetic field that is provided by helical coils surrounding the plasma. By monitoring the shape of the surface of the plasma and varying the electrical input to the coils, playing whack-a-mole as you put it with the shape of the plasma.

ITER has/will have a multitude of different control systems to control the inherently unruly plasma too. As I understand it, any tokamak will have to do something like that.

https://www.iter.org/newsline/-/3297

Poloidal field coils in ITER (and really any nuclear MCF machine that will ever be made in the future) integrate dynamic control. They operate on the microsecond timescale, which is expected to be sufficient for expected instabilities. You'd never move a vessel that fast. It's my understanding that instabilities aren't heavily affected by the vacuum wall's location anyway.

> Is there not a more active, not brute force approach for containing the plasma?

I think General Fusion's approach is interesting, they're compressing the plasma with a liquid metal wall.

This problem is clearly a fabulous use case for machine learning.

Yes but perhaps moreso in the more traditional ways. Stellarator geometries have several layers of optimization. You optimize for a last closed flux surface weighing MHD and turbulence models, then you optionally optimize nested flux surfaces, then you optimize single filament coils, then you optimized multi-filament coils. The computations are all too much, but more efficient optimizations buy performance for less computing.

I don't think so. Don't prematurely optimise. First thing to try is to hook up an RL algorithm and let her rip. If that doesn't work move on from there.

like I said above, we certainly hope so! It has been slow progress so far but applying modern ML / control techniques to tokamaks is one of the truly exciting applications of the current generation of AI in my opinion. Biased because this is literally what I do all day

Do you have a website or any papers on your work I could read?

I need to redo my website, just getting into the more public part of my PhD. Papers I would recommend from our collaboration on

control of normalized plasma pressure: https://papers.nips.cc/paper/2019/hash/7876acb66640bad41f1e1...

plasma profile transport modeling: https://iopscience.iop.org/article/10.1088/1741-4326/abe08d/...

hybrid dynamical modeling of gross plasma quantities: https://arxiv.org/abs/2006.12682

uncertainty quantification for plasma dynamics: https://arxiv.org/abs/2011.09588

It's still early days for this work and for us but we're looking at pushing reinforcement learning in methods and engineering to solve this problem

Do you have a github repo for the controller software? It will be fiberop to the sensors? And currentcontrolling by the ai?

Have you considered antagonistic training? One AI tries to destabilize the proces, the other trains not against a simulation, but against the destabilizing input and a succes-metric?

Reply on news.ycombinator.com