Automating the tedious calibration work of building quantum computers

Summary

Shai Machnes, CEO of Qruise, a company developing a machine-learning physicist, which is software that can automate the work of physicists building quantum computers, is interviewed by Yuval Boger. Shai and Yuval talk about the genesis of his company, the repetitive tasks that their software can automate, how smart the software can become and much more.

Read the full transcript on The Quantum Computing Report here

Transcript

Yuval Boger: Hello, Shai, and thank you for joining me today.

Shai Machnes: My pleasure.

Yuval: So who are you and what do you do?

Shai: I'm a physicist. I'm also the CEO of a company called Qruise, and what we do is we're building a machine learning physicist, and we're starting with essentially replacing ourselves, which is a machine learning quantum control physicist.

Yuval: So let me try to dive into that a little bit. A machine learning physicist: that sounds like there are tasks that physicists perform today manually, and you want to automate them. Is that about right?

Shai: Yeah. So if you think of primarily experimental physicists, and then think specifically about the relatively junior people in the lab, okay?

And now restrict this to things they do in software.

Yuval: Okay?

Shai: Because we're not building robots that will connect cables and actually build a device, at least not yet. So if you take junior people in the lab, they know how to measure certain things, they know how to calibrate certain things, they know how to do certain types of analytical computations.

And in the same way, if we think of how we train young physicists during their master's and PhD and the sort of things we teach them how to do, you can do the same with software. And this is what we're doing.

Yuval: Okay. So could you give me an example? In the realm of quantum, what is it that a junior person does on a regular basis that would be valuable to automate by software?

Shai: Right. So let's say we're trying to build a quantum computer. We're faced with a problem that either the qubits are not all exactly identical and not what we wanted to make them, or in certain situations, the qubits are identical because, let's say, they're trapped ions and all ions are the same, but the lasers are slightly different or the magnetic fields are slightly different.

So now you have to go qubit by qubit, and you need to measure all kinds of parameters about the qubits and about the control mechanism, and how the control mechanism can possibly distort what you want to do. And then you're trying to calibrate certain operations that the quantum computer will do.

The equivalent of, let's say, assembly instructions. With classical computers, you build a circuit that does a certain instruction, and you just replicate it a million times on the chip. We can't do that with quantum computers. It has to be ever so slightly different for each qubit. So now you have to go and do this for each qubit.

And this is a lot of work, and this is something one could automate. And then you can go further. Let's say I have a certain understanding of the physical system, and I want to measure a parameter of the physical system, say, how strongly two qubits are coupled to each other. You need to design an experiment, meaning decide which data you want to take from the device in order to determine this parameter.

This is something that a lot of the time physicists do manually with computers, kind of programming them manually, but we can teach computers how to figure that out. This is known as Bayesian experiment design. And then you can just point at the system and point at the model and say, "I want to measure that parameter," and the computer will figure out what experiments it needs to run so that it gets the data from which it can extract the value of the parameter.

Yuval: When would your system be used? Is that during the process of developing a quantum computer, or is that once I have a quantum computer that's live, say, on Braket or on Azure, and outside the operating windows I want to make sure it's calibrated correctly? When would your software be used?

Shai: Primarily during the development phase.

But quantum computers have this very annoying property that the very exact details of the qubits tend to change as a function of time. Certain frequencies drift, or things heat up a little bit, or atoms move on the chip, or whatever it is. There's a certain amount of maintenance that you have to do in order to make sure that things keep working at the really high fidelity you want them to.

So a bit of our software will also run essentially in real time. I think a good analogy for that would be disk fragmentation. We're going back to the days of spinning hard drives, before flash memory and stuff like that. You wanted files to be contiguous on the drive because then it's a lot quicker to read.

But if you read and write files randomly, the files stop being contiguous, and you need to clean up and do some housekeeping in order to make sure everything is right. And there is something like that that is necessary for quantum computers.

Yuval: Where does the machine learning part come into it? And related to that, do you expect your software to be as good as the human? Do you expect the software to be better than the human? Where does machine learning come into this, and what's the expectation of performance?

Shai: Okay. So the type of machine learning we're talking about is not the typical black-box, huge neural net model.

You know, people trained GPT-3. It has I don't know how many billions of parameters, and people are still writing papers to try and figure out how it's actually working and what it actually learned. This is a very different sort of machine learning. It is white box. It is physics-based.

You can use neural nets in certain places, like general function approximators, when you need a fudge factor somewhere. But the idea is that we're building devices. It can be quantum computers, but we're also looking at MRI, we're looking at silicon photonics and other stuff. We're building devices.

We need to understand in great detail what's going on inside, because we need to understand the sources of imperfection. Quantum computers these days have somewhere between, let's say, half a percent and one-tenth of a percent error rate per entangling operation. Now, if you have a half a percent error rate per entangling operation, that means after a few hundred operations, you just have noise.

So that's currently the limiting factor in terms of the capability of quantum computers. I'll give you a very simple example. IBM created a device that has 127 qubits, but the error rate is more than a tenth of a percent. Now, what does that mean? It means that they can't effectively use all 127 qubits, because just trying to entangle all the qubits, you have a less than fifty percent chance of doing this without an error.

Just the basic entanglement operation. Now, this is not a dig at IBM. IBM are doing amazing work, and they're state-of-the-art. But this is trying to explain where the field is at. Now, if we had 127 qubits like IBM has, but instead of, let's say, a tenth of a percent error, which we're not even at, let's say we had a one-in-a-million error, then already now quantum computers would be superior to the biggest supercomputers.

It's not the number of qubits. We have enough qubits. The qubits aren't good enough, and that's true for trapped ions and Rydberg atoms and NV centers and quantum dots in silicon and any other mechanism you want to think of. We just don't have good enough qubits yet.

Yuval: I understand the problem.

I'm trying to understand how good you believe your solution will be. Does it perform better than a human? Or does it just perform as well, but of course with much less hassle, and can do these repetitive tasks easily?

Shai: So, you know, there's a difference between the software as it is now and the software as it'll be in a few years.

But let's go back to the fundamental problem of quantum computing right now. There was a very famous paper by John Martinis, who was later the head of the Google quantum computing group, in 2014. It was called "Superconducting Qubits at the Threshold of Error Correction." They had a gate there with, I think, a 0.6% error rate.

That was 2014. Five years later, the Google supremacy result: the median error, again, 0.6%. Today, IBM's most recent 127-qubit device, the median error is slightly better, but not an order of magnitude better. We're making a lot of progress in the number of qubits. The rate at which we're improving the accuracy of the operations, the error rate, is not improving commensurately.

And the question is why? I believe the reason is that we don't have a very in-depth understanding of where the error comes from. This residual quarter percent error, or whatever, we don't fully understand where it comes from. The reason we don't understand is that we don't have a very, very detailed model of what's going on inside the quantum device.

And the reason we don't have a very detailed model is that you'd need to measure thousands of parameters in order to fully capture what's going on. And that's too much. That's too hard. And I think that once we have a machine learning physicist, we will be able to get these very, very accurate models of these very complex systems, and that will give us an understanding of what our current limiting factor is.

And once we understand what the limiting factor is, we can target the efforts in the next iteration of hardware better. So quantum computers will get better quicker, and eventually we will have useful, large-scale quantum computation. I think that with quantum computers, we're essentially bumping up against the limitations of what humans can do in terms of technology.

Now, of course, you can have an Apollo-style program or a CERN-style program. You put thousands of scientists on it, unlimited budgets. You can do that. And I think the big players, IBM and Amazon and others, are trying to go that route. You have a blank check, and you can just put in as much effort as you need.

But if we want quantum computers to be developed elsewhere, in academic labs, et cetera, we can't rely on huge teams and huge amounts of effort. And a machine learning physicist, it's a machine learning junior physicist, right? I'm not saying it's going to replace professors. It's going to replace young PhD students in the lab.

I'm sorry for the background noise, by the way; nothing I can do about it. If you give a lab an army of 100 PhD students that work 24/7, could they make progress faster? I think the answer is almost trivially yes. So this is what the software does. And maybe this year it's a master's student, next year it's a junior PhD student.

Maybe in 2030 it'll be a postdoc. But it'll take time.

Yuval: I think the company is about a year old.

Shai: Correct.

Yuval: Tell me the origin story. What caused you to start this company, and where are you today in terms of funding, how many people, where you're located, anything that you're willing to share?

Shai: Yeah. So the company started with three professors, Tommaso Calarco, Frank Wilhelm-Mauch, and Simone Montangero, and myself, and we're all working in control of quantum systems.

I'm the most junior of the team, and I've been doing control of quantum systems for a decade. So initially we did it ourselves, and then we did it through graduate students, and around 2000 we started feeling like the size of the activity in quantum was getting to the point where it could support a commercial company. So that's around the time we decided to create the company, and it was funded in late December of last year.

Now, we're a German company. We don't have a physical office. We're completely, 100% distributed. And the reason for that is it really helps with recruiting talent. The limiting factor in companies like ours is getting the right people. And no matter where you put the office, most people live far away from that office.

And by having a fully distributed company, we can get a lot of people, and a lot of good people. So we're 13 people today, and 16 or 17 people by the end of January. So things are going nicely. We recently got the EIC Transition Grant, so that's two and a half million, which is very nice.

We're really proud of that. We're involved in some research projects, and we're starting to work with commercial customers. What else? Oh, we meet once a month for a couple of days for brainstorming and planning and things like that, which is really, really important because it creates the social mesh.

We share meals, we share a beer after. It creates a social mesh on top of which you can start collaborating in a professional capacity.

Yuval: Do you feel that the software is particularly good for one type of qubit? Or does it support everything? I mean trapped ions and neutral atoms and superconducting and so on?

Shai: Yeah, we support basically every type of quantum hardware out there. We're working with the Bloch lab, for example, as part of a German research project, and we're doing Rydberg atoms over there. We support basically everything. But actually, it's not even limited to quantum.

And maybe going back to your question about the story of the company: when we started, we were thinking very much in the direction of quantum control. But it pretty quickly became apparent that the methodology and tools that we have can apply to other systems. So we're doing MRI now.

We're starting to do silicon photonics. If you think of a silicon photonics chip, for example, you have a lot of analog components on this device. And due to, let's say, manufacturing variabilities and stuff like that, they don't do exactly what you want them to do. Close, but not quite. And you need to understand exactly what's going wrong.

You need to build a detailed model of what's going on on the chip, and then you need to apply corrections so it works at the highest accuracy, which is exactly the same sort of problem we have with quantum computing. It's not based on the Schrödinger equation, though. We like to build a digital twin of the device that we're working with.

So if for quantum computing or quantum sensing the digital twin will be the Schrödinger equation, then for silicon photonics it will be propagation of EM fields and light and stuff like that. But it's the same basic principle. You have a system model, it's simulated with a digital twin, and then on top of that, you can do a lot of things.

We even once, for a VC in the seed round, created a digital twin of the solar system. We took data from NASA JPL on where the planets are and what speed they're going, et cetera, et cetera. And we used the same control algorithms to get a rocket from Earth to Mars. It's the same control algorithm that controls quantum states in a quantum computer.

As long as you have a digital twin and it's differentiable, so it has certain mathematical attributes, then you're good to go.

Yuval: So you could calibrate the moon. That's very cool. Well, so my last technical—

Shai: You know what you could do, actually? You can calibrate, for example, for a mistake in the mass of the moon.

Because if I can use my algorithms to realize that the frequency of the qubit is not what I wanted it to be, I can, by the same token, realize that the mass of the moon is not what I thought it is, and then correct for it. So in a sense, yes, you can correct for the mass of the moon.

Yuval: My last technical question.

It sounds like, for a customer to use this, it might require a lot of heavy lifting in terms of integration: creating the digital twin, connecting your system to the measurement methods and systems that we use, especially when building a quantum computer. Do you see it the same way, or is it easier than I describe?

Shai: I think there is heavy lifting, but it's not the customer's work, it's ours. When you get our software, you get a few dozen system models already baked in. So now it's more a question of playing Lego and connecting this and that. And if you happen to have a new system, something really unique, we'll help you define it in the system.

That's really fairly easy. Just to give you an example, it took us a week to do trapped ions from scratch. It's really not a lot of work. And in terms of integration with the experiment, we have a lot of experience because we've integrated with a lot of stuff. We have an API, we have documentation, and we can help.

So that's not really the pain point. The hard part is the real hard part, which is trying to understand what needs to be in the model in order to capture the behavior of the system. And that's where our machine learning runs up against its limit, right? Here you need a physicist to say, "I may have non-Markovian pink noise on the control line, and the readout lines may have some crosstalk effect, so please add this to the model."

This is something that our system doesn't know how to do on its own, and honestly, it'll be a while before it does.

Yuval: So Shai, if you could have dinner, or I guess in Germany a beer, with any of the quantum greats, dead or alive, who would that person be?

Shai: Hmm. Everett. There is one of the interpretations of quantum mechanics—I don't like the word "interpretation," but what's known as an interpretation of quantum mechanics—that is known as the relative state formulation, or the Everett formulation, or the many-worlds formulation.

Now, I don't like "many worlds" because it's a misnomer, and it's a really bad name and bad marketing. There are certain questions about the limits of what the many-worlds interpretation doesn't solve that I would love to discuss with Everett. And as far as I know, it's a pretty—well, I don't know if it's a sad story, but from a physicist's perspective, it's a sad story.

He came up with this brilliant thing and never got a faculty position. So he dropped out of physics and did I don't know what. But anyway, he's not alive anymore, but I would have loved to talk to him.

Yuval: So Shai, how can people get in touch with you to learn more about your work, and what kind of people are you particularly interested in hearing from?

Shai: I'm interested in hearing from anybody. So first, there are a few positions we need to fill in the company. If you have experience with trapped ions and quantum control, then we'd love to hear from you. But beyond this immediate need, anybody working at the intersection of machine learning and science, because the project of building a machine learning physicist is going to take a while.

This is not something you do in a year, or in three years, right? It's already useful now, but there's a long way to go. So anybody in the overlap of these two fields, I would love to talk to. And the way to contact me is just shai.machnes@qruise.com. As for how to spell my name, I'm sure you can find it in the description of the podcast.

And Qruise is with a Q, but I'm sure that's in the podcast description as well.

Yuval: Of course it's with a Q. You know, anything in quantum is with a Q. So Shai, thank you so much for joining me today.

Shai: My pleasure.