AI as an Accelerator for Fusion
So next up I have Dr. Richard J. But uh Dr. Butri is a theoretical plasma physicist and director of D3D lab at General Atomics. But Dr. But began his career in UKA where he worked on mast and uh jet for 16 years.
He's an elected fellow of Institute of Physics and American Physical Society and is recognized for his pioneering work on understanding MHD in Turk plasmas. So Dr. Richard But there we go. All right. And I'm delighted to see the US program represented by at least two Brits today for different parts of it. Um so today I want to talk a little bit about the fusion landscape. you
know, we we will talk about how we're using AI, how it can help the various different bits of Fusion, but but I think an important takeaway is that Fusion also provides a unique test bed for AI because of what some of the things I'll show you, the challenges we bring, which go a little bit beyond the normal applications. Um, you know, and the other point of course is if if Fusion is going to deliver in the 20 2030s, um, we're going to need AI to accelerate it and you'll see some of the ways it comes in here. So this talk is going to be a little bit of the landscape, the opportunity and of course what we can do in D3D. What we are doing already is and seeking to ramp up. Um we're talking about taking D3D as a major platform for AI. We have a lot of engagements already but can we take it to the next level and make this a major program at the facility. Um so we'll
start with the landscape. Um Steve has covered some of this. I'm going to go fairly rapidly over the introduction, but of course we've had major accomplishments in fusion. N has shown
that we can get more energy out than in. Steve showed Jet sustaining 70 megajoules of fusion energy achieved at jet. Um, terrific result. The Chinese facility is now going to a,000 seconds in high performance regimes, showing you you can sustain these things. And and
here on D3D, of course, we're doing a lot of fundamental science, but we're also pushing the boundaries. Here we push the density limit and that means you get more fusion and better power handling you know while still in a very high uh energy conf containment configuration. Um the fusion landscape uh it really is now at a burgeoning ecosystem. So you have the public
programs the national labs the national facilities um and they provide a terrific basis of expertise. You know we have worldleading experts in many fields materials experts at Oakidge computational people uh at Princeton and so on. Um and they've developed many of the foundations and of course the basis for the Eater facility which is now 85% constructed and people said why is fusion accelerating now? Well, it's kind of like we're 85% through building a power plantized reactor. It's time to be looking at the successes and improvements and so on and that's I think why the private sector is investing and why everybody is now looking at next steps. So our international partners are accelerating towards next step fusion devices. Some of these are funded. The step facility,
the Italian machine, the IT DTT. The Chinese are building a new facility that's going to be operational in the next few years called best that will be a long pulse superconducting high performance facility. And then there are other things in the pipeline with the Koreans, the Japanese uh and US objectives. CFS in the United States wants to build ARC. So we're really targeting this path to an aggressive path towards fusion energy now. And a
lot of it is based around the tokamat concept and you look internationally. But then the third element is the private sector that have made enormous investments now. So they've seen the benefit. People are voting with their dollars um to drive a more aggressive path and the fusion industry has this chart of you know the growth of the the sector. There's new machines coming online. Some of these machines you know
TAE who've got a great m machine that is demonstrating results. ST40 in the UK is now demonstrating world-class results, high performance plasmas and there's other stuff coming together with things like the Spark facility and Devons. Um, but a lot of this has been a science program uh and we need to also turn it into an energy program. We need the technology. So the question is how do we accelerate fusion? Now I think Steve's covered this but but there's two sides to this, right? We still need better plasma solutions. And Steve has made this point very clearly, right? Because the better plasma solution decreases the load on the plant. It eases the
technology challenge and it decreases the cost. But there's a range of technologies that also need to be established. We have to make the fusion fuel. We need nuclear hard materials. We need heating systems. We need very
efficient ways to extract the power. Um, and we need we need actual configurations that we can work out that can handle all the stresses. How do you integrate this into an engineering design? Um, so this is another part of the challenge that needs to be addressed and then integration between these two because of course the plasma solution does drive the challenge for the technology. So putting all of that together is each thing is a research project and putting it all together is a great challenge. So fusion needs AI to
accelerate the path and I I just want to give one example because this is what made makes me obsolete as a physicist. This is my field. Here's a tearing mode, an island. This is a tearing instability
that shortcircuits the energy confinement. It just gets out across the island and the whole thing can become unstable. And we spent many years understanding why these islands form in the plasma. You know, it's a combination of flows in the plasma, current gradients that drive them, turbulence plays a role, ion polarization currents. It's very subtle. And the bottom line is
you know we had a great productive research program understanding this island but sometimes they happen and sometimes they don't and we can't predict it and so Q uh machine learning this is what I find depressing so the scientists worked for 20 years worldleading experts and so on machine learning comes along and and we do some of our here we're doing hazard analysis this is fairly early work uh and it's pulling out now key parameters that are determining the trends and give us clues about the physics so on the one hand they're gener generating predictive uh better levels of prediction for the incident of these events, but they're also pulling out some of the physics variables in greater depth than we could as scientists. And so this this is one of the things that's been driving forward. There was the Egamman Coleman work in nature that Steve showed. So this these AI techniques are helping us discover the underlying physics um and they're they're giving us ideas of trends and dependencies that is very useful and they're also helping us predict these events so we can diagnose the plasma state and see when they're coming. So this is an important part of this using the AI to give you clues and advance the science to solve the science to connect to to codes and so on. Then
there's a bunch of other applications which we'll touch on. So just briefly here uh controlling the discharge in real time, deploying much more sophisticated analysis in real time, protecting the plant, de developing designs of discharge trajectories with things like digital twin. Diagnosis of a fusion power plant's a big challenge because you can't put many measurement systems in a power plant and the ones you do have to be neutron hard. So you need to be able to project the plasma state with very few measurements. Often measurements that are a little tangential to the thing you really want to know. Um and this is where AI comes in. Uh there's vast amounts of data here
and very diverse data. That's I think that's the important point. Um and we want to take that data and turn it into useful knowledge. You know, be able to create databases that allow us to understand the trends, maybe run codes, maybe do analysis. So curating data is
an important part of this. It's a datari-rich environment. Um, and model extraction. Steve's talked about this, you know, how we can connect uh how we can make the job easier for the supercomputers, how you can train on supercomputers, replace parts of the supercomputer train of events to to actually develop better models and much faster models. So, deeper, more datadriven resolutions is is the theme here. Um, but Fusion also provides a unique testing ground for AI. So you
know Steve has talked about this there's an opportunity to stress test and augment the models you know interpret understand and predict but the thing I wanted to point out here was there are really beyond supercomput challenges and if you look at a fusion plasma there's a lot of stuff going on you've got a collisionless uh fluid in the magnetic fluid in the core but going right to the edge you've got a very cold plasma uh that's collisional and there's a there's there's things going on you got heat coming out Here you can have nonlinear explosive events. You have turbulence impacting the things. You've got island structures. You've got stochastic regions. And then as you sweep the power out down to this power handling regime, you've got cold recombinative plasmas. So there's a vast range of scales going from the micro to the macro and they interact with each other and affect each other. You know, and you're looking at
order of magnitude ranges up to 10 to the 10. So there's a time dependence and a space dependence that this you know you can do codes to do little bits of this but putting it all together is actually beyond what we can do now. Um so this is something I think where you're going to need the AI to bring it all together and accelerate it. Um there are other examples I think there's people we were talking to last night doing fusion materials that you can use these to go from you know the micro how the the latises and so on to the bulk properties and facility design. You looked at how that all of that plant integrates putting that again together into a model to design an optimal facility and work out the trade-offs is something that's going to be challenging. So, so I think these are things where go beyond what you can just do with a computational technique uh that that's going to need AI to bring it together. So, how can we synthesize
things? Um, and and then finally, of course, D3 D3D fusion brings a huge diversity of data and real world devices are a little different and they can help push the boundaries of the these techniques. So, it's it's a it's a novel test bed and we'll get into some of that. So, Fusion engages a very wide range of AI challenges. Steve has
explained the role of the TOKAC here, but the key point I wanted to make here that there has been tremendous progress on the TOKAC. You can see here over the years that the devices have got very close to the burning plasma conditions. And this is why we talk about the TOKAC because it's the one that's going to get there first. There might be better ideas
like stellarators or some of the wild fusion concepts that might be better in the end, but TOKAX are close and they're going to be in the zone where you demonstrate the techniques soon. Now they are more soon. And so this is the test bed you want to use if you want to develop AI or other technologies for fusion. So I want to now introduce D3D.
So D3D is a Department of Energy national user facility. We're one of 28 office of science national user facilities. We're the only one that's run by a private company, but we have an open shared leadership program. So, it's a worldleading platform that's focused on innovation. That's a great tagline.
It's a bit bland. Okay. But the points of the facility are we we have high flexibility and measurement and we can rapidly change out parts of our machine. We we're opening it up every year to put new stuff in. We are big enough to be
relevant, but we're small enough to be fast. We'll have people come in and tinker on the machine every 20 minutes because we can do that. Um, we're a user facility that is served for the nation. We have 700 users, about 100 institutions, leading labs like Princeton, they have a big collaboration with us, universities, private companies, and we have an open user model. So, we'll get into some of this. So, it's a live data producing facility at the cutting edge. We have experts in every field working at the facility. So
if you're a user, you can talk to them. Um and um and we're innovating new technologies and approaches including AI techniques. But as a national user facility, we are provided free as a national resource. Our job is to serve
the national interest. So a little bit about D3D. It's enormously flexible device. We can pull out all these kinds
of different shapes. There's a new one here that's everyone's getting very excited about at the moment called negative triangularity where you take our normal shape and invert it. And it turns out that's better in some ways. We can control the whole plasma. We can inject heat and current and momentum in different proportions where we want it in the plasma because we have a lot of different heating systems. And so we can
change the design of the internal configuration of the plasma where the current and rotation uh and and and heat is. Uh we can test many different techniques. We're looking at like fueling technology, impurity injections, different materials, different field configurations, different RF technologies. Um, and we have an advanced control system that gives us a robust platform for us to try out different techniques, dialect configurations that can stress test things. So to dis this is really aimed
at discovering the science and techniques of fusion. Uh, but maybe where D3D shines is it's has the world's leading diagnostic set. So we have about over 80 measurement systems uh 50 different underlying techniques provided by many groups from across the United States. Um so they measure almost every property of the plasma. So it's highly heteroggenic data source. You can see
here this this plot is a basically just a photograph of the inside of the machine stretched out. And every single one of these ports has one or more systems that's provided say coherence imaging by Livermore or or or beam emission by UW. Um you know each one of these is a state-of-the-art technology that is measuring something some remote property of the plasma. Um and so we can
understand a lot. Um so we have over 80 systems and we're we're characterizing every aspect of the plasma behavior. So if you introduce a new technique, a new technology, uh you want to understand what's going on as a result of that, we can diagnose that uh and understand the processes and then we have many theory groups engaged in the facility as well who are then using using all of this data to test the models and so on. And machine learning and AI is an increasing part of this. But for us the why is critical because you have to understand enough that if you're going to project that fusion reactor you need to understand that your projection with confidence. Um just a quick tour of some
of the range of the facility and the underlying point here is that there are many common processes uh and questions between all of these different fusion concepts. They're all plasmas. They're all MHD fluids. They all need certain materials and so on. So there are
differences of course, but there's a lot more in common. And so D3D is being used to look at a wide range of things. So for example, here we're doing an experiment which is halfway through executed now to test compression heating. That's General Fusion's crazy
idea. They're going to do it with liquid lead being pumped down to compress the plasma, but we're just doing it with our coils, and we can test some principles for them. We um we we're testing advanced materials. So here we're looking at materials that were used for spacecraft entry into Jupiter where the material didn't behave as planned. So we're exposing it to D3D plasmas which is the photograph here and learning things about the material models. Um we're looking at fundamental physics. We
work with the discovery plasma physics community. So here we're looking at types of reconnection that's associated with uh solar flares. Um here we're looking at wave particle interactions which is associated with the magnetosphere. So, you know, how space weather impacts uh the the planet. Um, and uh here we're looking at organic molecule formation in plasmas, which I like to say is the D3D discovers life or discovers the origins of life or so on, but it's it's putting a little step in the ingredient as to how you form organic molecules, uh, how that might happen in space. And so, the diagnostics
we have give us a lot of insights, but it's a very broad capability that we have with the machine. And then a final side I wanted to talk about materials because this is something we can do very well. We have this facility you know all the pete comes out down in this bottom zone here but we can bring in samples every discharge. So we can change out
put a sample in and expose it to plasma and see what it does when you put 10 megawatts per square meter onto it. And so this is highly diagnosed. It's very flexible. And we have a pretty strong program now of of bringing in different types of materials for evaluation. Uh as we're moving towards this technology program, we can also change out whole arrays of tiles in the machine and we do routinely do that. But we're using this
to explore see Microsoft is the font's gone. Okay. But you could we're using this to explore a range of different types of materials like high temperature ceramics, dispersoids, um can't remember that one, plasma spray technology, liquid metal techniques. So these are things where we've had samples into the machine in the last year looking at some of these things trying to advance a bit and and get some insights. So that's that's kind of D3D.
Just a couple of slides on how this works as a user model. How am I doing for time? Right. Okay. Good. I'll slow down. Ken was nervous. Uh, so, so D3D works as an ecosystem and we love this. This is it's a team and
people have to work together and help each other out because are complicated and hard and you need many different things to come together to make an experiment or make a fusion device. Um, so we are a DOE owned facility which means we are provided free. Uh, there's no charge for using D3D. You become a user. Um and and you the facility
resources are provided you know teams help each other out people providing systems will help somebody else with an experiment u we provide runtime we provide data training office space and so on and if you bring something that needs specific implementation you know you need a technician support you need computational staff or whatever DOE we have to go to DOE for the money for that but DOE will typically fund that just as they fund our present users when they win a funding award from DOE it's a team model so typically a user brings a capability could be a measurement system it could be an AI system um but you you provide it to each other you're part of the team we have worldleading experts I said this already I think that that are pulling together so there is an example they tell for ST40 which is a fusion machine uh in the UK uh they had a problem with a measurement system that they were struggling with they they got an expert actually from the British fusion lab over um this was facilitated by the US program somehow. But but that person was able to give them advice within half an hour that that would have saved them months of research because they knew about the technique. There's a depth of expertise in all of the techniques in the government funded programs at Princeton, at D3D and so on that you can just ask somebody a question, you get an answer that you would have struggled with or not solved.
Um and it's a shared leadership model if you're engaged in a non-proprietary way which we'd recommend. So basically we have program areas you know materials transport whatever it is RF technologies and the groups get together and they collectively discuss uh uh discuss plans and how to take the field forward. So it's a meritocracy uh and it's a collaborative experience but of course it's not just bottom up. This is tension by we have strategic goals for the facility which we work with department of energy to agree uh the leadership tries to cascade those down. there's a
bottom up and a top down and we have a user board overseeing us. So this is a huge resource that's doing cutting edge science. Uh and one of the things we have done recently is reorientate our approach for private sector engagement. You know, we did we did my deputy David Pace did his MBA on on private sector fusion engagement and we looked at what the fu what fusion companies need and and there's a list of things here but solving technical challenges was the number one issue from surveying 22 fusion companies and so we reoriented our program to target technology goals and align with the private sector. And the important point here is the program goals are to accelerate fusion technology. So when a user like Microsoft or or Nextstep Fusion or whoever comes in and wants to test something, that's already part of our goals and that's why we can allocate our resources to it. Um and so we we have a
non-proprietary user agreement that allows you to protect your IP. You know, if you've got some clever software or clever measurement system or a new material, you don't have to disclose what that's in. You only have to have an intent to disclose what you get from D3D. So if a D3D system measures something, you can't keep that secret or patent it or something, right? We provide support and it's a partnership approach and this has really been going very well. You know, over the last few years as all these US strategic processes happened, we developed this legal framework. Uh we developed our plans and now we have something like 14 private companies involved here including a lot of the big fusion players um and Nvidia as well. Uh and so
this this model really provides full and equal access to the program. You're not committed when you sign up. It doesn't cost anything and you're not committed and if it's not interesting, you just don't participate. Um but if you do engage, you get to talk to the experts, haggle out what you want to do and and win some program uh support. Um and you're competing against things. So for us, this is really important. We have to
use these national facilities, national labs and so on to help the private sector accelerate. And there's a balance here. It's not that we defer completely to the private sector, but the private sector is driving a lot of growth. So,
it's how do we work to solve their problems and help them get there sooner. Okay. So, I'm just going to introduce this and then Dave Humphre is going to pick up the details. So, so we this is just a quick uh tour of what we're doing. So um so D3D is pioneering digital technologies in fusion with private sector. So you know a part of
this is a digital twin of the machine uh that you can replicate design things and and use that to simulate things you might do on the machine. Uh but this is sort of one part that then connects to the real-time control. You can you bring in more advanced uh control techniques based on machine learning based on representing uh advanced simulations and so on. uh so this also connects into uh you know curating the data and feeding it off to big supercomputing facilities and we have programs where we connect D3D direct to big supercomputers the moment we get intershot analysis but of course as you go to surrogate models you might want to be bringing that in for real-time control so data is a huge amount of this of course you have to process and curate the data I'm going to talk about that in a minute um and then you know actually doing simulation to design your discharges is your better configurations or or design a power plant as well. Uh and we kind of do both of those things uh at at the facility. But the point is if you're going to build a fusion power plant, you going to want a digital twin first. And if you
want a digital twin, you need some hardware to validate it against. And that's where we come in. So here's just an example I think of some of some of the work you know where machine learning is basically using a cost function to work out the optimum trajectory and then that's feeding back into into finding the optimal path in real time with the black line on D3D. But I think what I wanted to say was just getting started. Um we do have some strong engagement now with private industry and we would love for Microsoft to get involved if you're interested. We think it would be a great experience. Um we have been selected for
the office of science pathfinder program. This was another white house initiative. Um and we're very interested to ramp up what we're doing and really take this to the next level and make this a major program at the facility uh and put in a bit more underlying infrastructure which we think we can persuade DOE to resource if we have good partnership. And just as one step along the way of this um you know we have now implemented a fusion data platform. This is live. It's serving D3D data. This was
based actually on some of the software for CERN for providing uh providing the data. But then we've adapted it for fusion and D3D was key part of the developing the techniques for the adaption to the fusion. We have other machines coming in now and then this will come in um to with a with a supercomput techniques as well. Um but
but I mean so the point here is to serve data cur in curated ways in large amounts very rapidly and the point we'd make is um large new bigger facilities in the future will have more data but they will have less diverse data so D3D is at a point where we have can measure everything in many ways and so this is the sort of unique opportunity to develop the approach with the diversity of data we have and so this is the conclusion We would say fusion is a great testing ground for AI. AI it's a great accelerator for fusion but the other way around. Um and D3D is a platform to do this. You know we're highly flexible. We
can measure. We have an open user model. Your goals are our goals. That's the point. Um so this is the team and come join us. Thank you. [Applause]
2025-05-16 00:54
