3. May 2023
The head of the Nanomagnetism and Spintronics research group, Dr. Vojtěch Uhlíř, can be proud of his success – together with foreign partners he will be working on the METASPIN project, which has received support under the European EIC Pathfinder Open call. The ambitious project, coordinated by the University of Paris-Saclay, aims to create a revolutionary technology for low-energy artificial synapses using spintronic nanodevices. This technology could solve a major problem in artificial intelligence applications – catastrophic forgetting. The research group at CEITEC BUT will focus on the modification of prototype circuits using antiferromagnetic materials, which have an unusual internal structure and do not exhibit magnetic properties externally. The team is thus faced with a challenging task – they must find a way to read and manipulate the magnetic orientation of these materials.
How does your group contribute to the project?
We contribute mainly in terms of materials. We want to modify the prototype scheme with a new class of materials that could improve the functioning of artificial synapses. But of course, we don't know if it will work, for example, from a physics point of view, if the mechanisms and materials will be applicable. So, it's an application-oriented project to some extent, but we can play around with how we end up contributing. However, if by chance we fail, we still have the option of using traditional ferromagnetic materials to emulate synaptic behaviour, i.e., to mimic the workings of the connections between nerve cells.
What class of materials do you have in mind?
In our case, we are talking about so-called antiferromagnetic materials, which are special because they do not exhibit any magnetic properties externally. To illustrate, imagine you have a conventional bar magnet, and you scatter iron filings around. These magnetise around it and form a pattern. That doesn't work with antiferromagnetic materials. If you throw iron filings in, they'll behave the same way around a piece of plastic or wood. And that's because the internal structure is arranged in such a way that the individual magnetic moments actually cancel each other out.
Anything else that makes these materials unique?
They have a specific arrangement. It's not like, say, aluminium, which has magnetic moments on the atoms arranged randomly. Here, they have a precise structure. Which can be exploited if we can figure out how to locally read the orientation of these moments. For some materials, all we can tell so far is what their direction is, if they're rotated, say, 90 degrees relative to each other. So that's the first challenge we have. And the second is how to control that rotation. Because since they're not emitting any external magnetic field, there's nothing to grab onto. Normally, we would apply an external magnetic field to the ferromagnet. The antiferromagnet won't care to the first approximation.
Are there any advantages to this behaviour?
Actually, many advantages. If you had a component based on an antiferromagnetic material, you would not have a problem with external magnetic fields, and therefore, perhaps a loss of information or unwanted manipulation. But on the other hand, it's very challenging to figure out how to manipulate the material in a purposeful way.
Do you have any ideas how to do that?
In the last five or six years, there has been a fairly strong development of an approach that involves manipulating the orientation of antiferromagnets using electric currents. Depending on which direction we send the electric current into the crystal, we're able to rotate the internal magnetic moments. This approach, of course, needs to be verified for the materials that would be of interest to us. But beyond that, we need to come up with an additional mechanism to either make it easier or harder to switch or change the orientation of the antiferromagnetic moments. And that simplification or complication is already related to neuromorphic behavior, that is, how the recording of information in our brain works and how we distinguish which information is less important or more important. Which brings us to the heart of the METASPIN project.
So what is the aim of the project?
The inspiration is the brain, specifically its synaptic behaviour. Repetition is important for learning. The more often we do something, the easier it gets until it's automated. In a similar way, we're trying to emulate a memory effect in a component. We can imagine that if I had a wire that I was sending current through, I wouldn't expect that the more I sent current into it, the more its properties would change. But there are materials or physical mechanisms that make this possible. The more often the signal goes in, the more the basic resistance of the wire will change. So it may be easier or harder to send a current there depending on what mechanism we use.
What could such a principle be used for?
Right now it's being used at the startup level, where some simpler specific task needs to be automated.
How can we imagine this?
For example, you need to process documents where you need to find something specific –some text, an image. If a human is doing that, they look at it and say, yeah, that's what we're looking for and sort it by that character, for example. For a standard software approach, this task is challenging because we have to find the pattern in the document first, and that takes energy. For example, we have to ask if the pixel is black or white and then look for some other pattern in it. So it's quite a hard task for a computer, but it's not hard for a human. If I walk into a full lecture room and you ask me if my student Christopher is there, I'll look and in a few seconds I'll say yes, no. And it doesn't even cost me much energy because I know what Christopher looks like. I've seen him several times and it's fixed in my synapses. It's like looking for a pattern in that particular document or image. So that's in a nutshell how neuromorphic behavior works, and what we're trying to do in the components.
With memory comes forgetting, is this a process you want to work with too?
Yes. We'd like the component to learn multiple tasks at once. If you teach it to look up, say, three students in a lecture room, and then you move to another course where other students are attending, and you start teaching the component what they look like, so that it doesn't forget the image of the three students in the original lecture room at the same time. Which is a task our brains can handle. Of course, if you didn't see them for a long time, you'd gradually forget them. But with a hardware component, the forgetting process happens very quickly, which we'd like to change. We need to figure out how to slow it down or modify it so that we can still use the previous task. We can't make forgetting impossible.
Do you know what means of reducing memory loss you're going to try?
Again, we look to the human brain for inspiration. When you have some additional sensation – sound and image – while you are learning, you are more likely to remember more. Or if it is connected to an emotion, for example, the speaker used an analogy that is close to your heart or told a good joke. That's something that will help you remember it. That's what we'd like to try. In addition to modifying the material with one type of stimulus, we need to give something else to the speaker that will set the system up to remember it better or worse.
Which is actually a value-add, because it doesn't work that easily in humans.
It's like a humorous statement commonly associated with Jára Cimrman, a fictional character from Czech culture. The joke is that the speaker tells students what they should remember as important and what they should forget as unimportant, but the punchline is that typically it ends up the other way round (laughs). But we can say to the component, here this is important. If you write with shift – now it's going to be uppercase, now it's going to be lowercase – somehow you can do that. But the mechanism, how do we make it just be "more significantly" written or "less significantly" written. That's something that we're going to explore and that we want to apply to antiferromagnetic material.
We're talking very abstractly now, how can I imagine your work in concrete terms?
We've been discussing the concept of the project so far. There's a broad consortium of people working on it. Some are making materials, some are mathematicians who are thinking about the algorithm, how to implement it for the material. Others are companies that are involved in, for example, developing planar technologies, preparing thin films, making chips, etc. And we are specifically studying magnetic configurations in small-scale magnetic materials. By being able to image them, we can test their control by different stimuli. So we're a long way from teaching the material anything.
So you work in a lab?
Yes, I'm working in a lab. Our work consists of preparing the material, structuring it into some interesting shape, modifying the material. Then we connect the contacts, make electrical measurements, using different types of microscopy from optical to electron. But if you're asking what exactly I do in the lab, unfortunately I do almost nothing. I try to coordinate it all, see where, what's "burning" and usually I just give some advice and then I have to make phone calls again to arrange stuff, then I also teach, establish new collaborations... So, although I don't like it, I'm moving more towards this kind of manager work.
I'm finding it stressful with all the different activities, but you give a very calm impression. It's like you're handling everything with grace.
It just seems that way (laughs). I think that's the only way to handle it and not go crazy. When we were just talking about brain functions and how we think, it's scientifically proven that a person is not able to switch after five minutes to a completely different topic. And it often takes me a while to "pull" what I need from that long-term memory to the operational one. I don't do things that I could easily automate. I have to concentrate every time, which is exhausting.
And how do you recharge the batteries?
I play sports, I garden. Preferably alone. I need something to do where I can switch off. Ideally some kind of automated activity. That's important for the apps, too. If we can teach the device to automate certain tasks, it saves us energy. It's like our brain. If every time you brush your teeth in the morning, you have to think about it. Oh, the toothbrush, what I'm gonna do with it and all that. If you have the pathways in that component, it's done with minimal energy.
You said that the antiferromagnetic materials you're interested in now have no magnetic field. Is there any way to create one? To make them look like ferromagnets?
You're referring to a principle. For example, in one class of material, we can induce a phase change from antiferromagnetic to ferromagnetic. You can imagine it like heating ice. You turn it into water for a while and then you cool it back down to ice. We're able to do that in the solid phase in one material. It happens near room temperature, so it can be interesting in that we can locally heat the material. Maybe with a laser. And somehow modify it in the ferromagnetic phase and then cool it down again.
And would that be advantageous in any way?
Maybe it could help with the controlled manipulation of antiferromagnetic states. As I said, we don't know yet if we'll even be able to turn the magnetic moments in the antiferromagnet as we need to. So these are different approaches that we can try as plan A, B, C, D. So right now I can't tell you if it's going to work or not. But it was a very good question.
What will be the final product of the METASPIN project?
A chip that can emulate the behaviour we have been discussing. It is then up to the companies involved in the project to come up with proposals for applications of our product. They can be used in recognition of multiple types of different objects, in image analysis where you need to process large amounts of data.
You mean like security camera systems at airports?
For example, but you can really think of more applications than I can. I don't dare say what engineers will end up using it for. I'm happy to be surprised.