Hello, and welcome to the Physics World Weekly podcast. I'm Hamish Johnston. This week, we'll be celebrating the winners of the Physics World breakthrough of the year award for 2024. This episode is supported by the journal Reports on Progress in Physics, which offers unparalleled visibility for your groundbreaking research. This year's award is all about error correction in quantum computing, and we're honoring 2 independent teams.
I've spoken to the lead researchers of both groups, and we're presenting those conversations in 2 different episodes of the podcast. So it's a real bonus this week, not one, but 2 weekly podcasts for your listening pleasure. In this podcast, I'm in conversation with Google's Hartmut Nevin, who led a team that has made a major breakthrough in implementing quantum error correction in a processor that uses superconducting qubits.
In a separate episode, I chat with Mikhail Lukin and Dolev Blufstein at Harvard University, who, along with colleagues, have implemented quantum error correction on an array of trapped atomic cubits. In principle, quantum computers can solve some problems that cannot be computed on conventional processors.
However, the quantum processors available today are very susceptible to disruption by environmental noise, and this destroys the delicate quantum states that are used to store and process information. When quantum computing was first proposed, some physicists thought that this problem was insurmountable. But thanks to the development of quantum error correction, practical quantum computers that can solve useful problems could soon be a reality.
Quantum error correction works by distributing 1 quantum bit of information, called a logical cubit, across several different physical cubits, such as superconducting circuits. In principle, the robustness of a logical cubit should be improved by increasing the number of physical cubits. But there's a problem. Boosting the number of physical cubits itself introduces errors, and therefore, creating an optimal quantum error correction system is no easy task.
This year, we've awarded one half of the 20 24 breakthrough of the year award to Hartmut Niven and colleagues at Google Quantum AI and their collaborators, and that's for implementing quantum error correction below the surface code threshold in a superconducting chip. For the first time, exponential error suppression in a logical cubit has been achieved as the number of physical cubits increases. And to chat about this achievement, Hartmut joins me down the line from California. Hello.
Welcome to the podcast, and congratulations on your team's achievement. Thank you for having me. So Hartmut, I suppose first things first, with with my questions. What is quantum error correction, and why is it needed? A quantum error correction is a necessary technology that allows you to scale up to large quantum computers with many cubits that can participate in many algorithmic steps.
And and the reason is because each individual qubit in a quantum computer, at least a a quantum computer that exists today, is is very noisy or subject to failure. So you you sort of have to club them together to to get one good cubit, cubit. Is that how it works? Yes. The way how I like to think about it, people often say quantum information is very fragile. We need to protect it. I like to think about it a little bit differently. Quantum information is very contagious.
Qubits like to talk to each other, and they also like to talk to Qubits outside of our chip, outside of our control, and that leads to leaking information, leaking out of the processor, and we need to prevent it. So quantum error correction is really a set of technologies or a technology that allows us to control all the information necessary for a quantum computation. I see. Okay. And and in the work that you've done with the Willow processor, how how have you done that? How have
you dealt with this leakage of information? How have you made sure that, the information is where you want it to be when you do your quantum calculation. So, quantum error correction, like classical error correction, draws on the time tested principle in engineering. That if you want to make something more stable, you introduce redundancy. Understand this is, a physics world audience, so maybe this example is too simple. But,
often say, hey. If you want to fly, let's say, from Germany here to LA, if you have an airplane with 1 engine, that will work. But 2 engines is safer. And if you have 4 engines, it's yet better because if one of them fails, you still easily make it over. We use the same principle in quantum error correction. So we want to represent 1 logical qubit or the information contained in 1 logical qubit.
So how we do this is we orchestrate a set of physical cubits, a little array of, let's say, 3 by 3 or 5 by 5 or 7 by 7 physical cubits that make one better protected logical cubit. To our delight, what we were able, to demonstrate is that as we went to larger arrays of physical qubits,
the error rate was reduced. So what we were able to do as we went from codistance 3 to 5 to 7, each time as we increase the codistance, the error rates were reduced by a factor of 2, effectively leading to an exponential reduction in error rate and therefore creating, the most convincing prototype of a logical qubit till to date. I see. And is that something that surprised you when when you set out to do this research? Were were you expecting to see that that exponential effect?
No. Theory had predicted this. It wasn't, reduced to practice yet. So we could actually, predict this, quite well. What is important if you want to achieve such a result is your it's a system engineering challenge. So it's not good enough if just your single qubit gates are very good or just your 2 qubit gates are very good. Your state preparation, your readout, all components have to be, very well, engineered and have to be what is called
below threshold. See, this only works, or if you want to have more cubits but less error. This only works if your cubits have achieved a certain basic quality, and that is known as the field as being below threshold. So if all components of the system are reasonably good, then you can orchestrate them into something really very good. That is essentially how quantum error correction works.
I see. And and and so the system that you have available at the moment, would you describe it as a sort of a proof of principle system? Or are you able to to actually use it to solve practical, computing problems, and and even problems that can't be solved easily by a a conventional classical computer. So the quantum error correction demonstration made just a single good logical qubit. Of course, a single logical qubit is not
good enough to run any interesting algorithm. You will need, many of them. And also, you need it yet better. We achieved roughly a 1 in a 1000 error rate, which means that then you can run about a 1,000 operations in your algorithm. Because the way you can think about error rates, if it's, let's say, 1 in a 1000 or 1 in a 1000000, what it means, you have your cubits, you apply your, gates, and then you have the 1 in a 1000 or 1 in a 1000000 chance that you
crash your machine. You get a blue screen and you have to restart your computation. So 1 in a 1000 error rate means we can really run algorithms with about a 1,000 gates. So this gives you a certain limit of complexity. Many of the famous algorithms, for, quantum simulations or factoring large numbers, they often need way more, gates than a 1,000 and need 1,000,000 or even 1,000,000,000. So, therefore, we have to, make
yet lower error rate logical cubits. And we can, of course, do this, rather easily in a way because once you have an exponential reduction, you just ask for this algorithm what error rate do I need. And then you go to the appropriate code distance, which, let's say if you want something really low, like 1 in 10,000,000,000 error rate, that's something we want to eventually achieve. Then you go to a code distance of 21 or 23, or if you improve the overall hardware, maybe we can do it with 19
or 17. Meaning, then you have a 19 by 19 array of physical cubits that make one very good, very low logical error rate, cubit. And then, of course and you don't need just one logical cubit, but you will need, let's say, a1000 of them. And that is since the endpoint of our development roadmap. Our roadmap calls for it consists of 6 salient milestones.
And the last milestone, at least for now on this roadmap, calls for building a 1,000,000 physical qubit, quantum processor, which would, with today's methods, translate into about a 1,000 highly protected logical qubits. Now you have a system that can confidently generate commercial value. I see. And and can you talk a bit about the the the physical hardware? The am I right in thinking that you're you're using, superconducting qubits. Is that right?
Yes. That's that's correct. There's a different ways how you can represent a quantum bit, a cubit. Essentially, any quantum mechanical two system state can be used. We use superconducting cubits. So, anybody here on the audience who has let's say when they were teenager built a little radio out of an electronics kit, they may be familiar with LC circuits or electrical oscillator where you can sync the the electrons slash back and forth between a capacitor and an inductor.
Our cubits are basically this LC, circuits, but they are superconducting, and live at very low temperature below the transition temperature of the superconductor. So they essentially implement a quantum mechanical harmonic oscillator. And people may remember from a 1st semester quantum physics, class that in a quantum mechanical oscillator, you have discretized energy levels. And we use the lowest energy level and the first excited energy level as our logical zero and logical one.
I see. And and how how do you connect up these, these superconducting oscillators? How exactly do you get communication and coordination so that you can do your error correction and, ultimately, computations. So so there are different ways. Once you have your, I'll see oscillators, your, your cubits. Maybe one piece I didn't, say, that's still important. I told you we essentially, implement cubits as quantum mechanical oscillators.
And I told you that we use the lowest energy level and see, the next highest energy level as our 0 and 1. We do one more piece because in a harmonic, quantum mechanical oscillator, then there will also be a level 2, an energy level 3, and so on. And they all have exactly, the same energy distance, hbar, times omega.
So that is not so good because, if I have my qubit, let's say, in the first excited state and I send it a pile with with just this energy difference, and I can exactly make sure it goes back to my 0. It might also go up to, level 2, and we don't want this because this is outside of the code space. So we put one more ingredient in, and that is the Josephson junction. And the Josephson junction is a nonlinear circuit element, and it makes our oscillator slightly un harmonic.
What this means is now the energy levels are not equidistant anymore, but the first separation between 0 and 1 is a little bit larger energy difference than between 12, and that is still larger than 23. So you get this letter of shrinking, distance energy levels. And that's quite useful because now we can send in control palaces that will only cause transitions between the lowest, two energy levels and the others out of
code space are not involved. So that really completes, the superconducting qubit. Now once I have a superconducting qubit, how can I couple them? There there are various ways how you can, couple them. There are capacitive couplings, inductive couplings. You can use little, qubits, in between that act as a coupler. You they just have to get into interaction. So you you can maybe think of the mechanical analog, you know, pendulums, and, you put a little spring between them,
and then they feel each other. That is an interaction between, 2 oscillators. So there are multiple design choices you have. And what's the exact, best choice for coupling is that is actually still a bit of matter of research. I see. And and you mentioned that, I suppose, your ultimate goal is to get a a quantum processor with that offers about a 1,000, logical cubits. And so so that would require many, many more actual physical superconducting
cubits. Is is that the sort of thing that that can be miniaturized onto onto a chip? Or or would that have to take up a, let's say, an entire lab, at a university? So, roughly speaking, our our cubits are not super small. Actually, if I were to give you a chip, which has an array of cubits and you squint, you can actually see, the individual cubits. They are a little bit smaller than a square millimeter. So, essentially, out of these components, we would make larger and larger arrays.
But currently, a little bit as opposed to classical, CMOS, technologies, we are not aiming or it's not a priority for us to make the, cubits smaller and smaller because, we actually like that they have the size because then it's easier to control them with, microwave pulses, so it's easier to read them out. So we would just have to make,
sufficiently large chips. Eventually, you have to, join multiple chips to make a chip that has the surface area to, cover, let's say, a 1,000 or even a 1000000, physical cubits. I see. And and so where are you in in the development, of that at the moment? Have you have you managed to to create, a system of integrated physical cubits that's large enough to give you a 1,000 logical cubits? Or is that something that's down the road a bit further?
So I mentioned the road map that, the Google Quantum AI team published. And this road map consists of 6 milestones. So the first milestone was we achieved it, in 2019. It was showing for the very first time that a quantum processor could compute a task in minutes. That's a zen fastest supercomputer would have needed 10000 years to do.
The second milestone we also achieved already was similar to the current experiment, was to show that as we go from code distance 3 to 5, the error rate comes down. But that milestone was defined as just being the break even point. So, yes, the error rate did come down, but just by a hair, 4%. So that was not that impressive yet. So the the current, result, which we refer to as a yard stone, is actually between 2 big milestone that improved, on milestone 2 by it's not 4% anymore.
It's by a factor of 2 that the error rate came down and it came down twice from, again, code is 3 to 5 to 7. So milestone 3, which is sort of the midpoint of our road map, will be a single very good logical qubit with a 1 in a 1000000 logical error rate. And then from there, we, now maybe I mentioned the remaining milestones. Milestone 4 is then having several, logical cubits of about that quality and to have a gate set, a universal set of gates, gate operations between those logical cubits.
As we then scale up through milestones, 56, we get a 100,000 or even a1000000 physical cubits, allowing us to make, more and more logical cubits, let's say, with current technologies or current estimates would be a 100 or 1000. Most likely by the time we reach those milestones, quantum error correction technologies have improved further. Our hardware has improved further, and we get a few more logical cubits out of a given set of or a given number of physical cubits.
I see. Okay. And do and do you have any any sort of feeling for, the the timescale, you know, when when you will get to a situation where you can you you can start connecting up these very good logical cubits. Is that something that is that like a 5 year thing or a 10 year plan? More about 5 years. We expect that we will have feature complete quantum computer with at least a 100 logical cubits, hopefully more like 1,000 by the end of this decade.
So, I've been asked this by reporters before, oh, is quantum computing like nuclear fusion, which famously is people quit, 20 years out. Quantum computing is not like this. We publish the road map and we pretty much knock out the milestones as clockwork. So we are making good progress. Of course, it's a very ambitious road map and could we be delayed by a year or 2? Absolutely, that could happen, but so far, so good. We have pretty much stayed to true to the predicted timeline.
I see. And and when you get to a 1,000 cubits, logical cubits, I mean, I would have thought that there were practical things that you could do with a 1,000 cubits? You you know, may maybe not solving universal problems, but are there are there specific problems, maybe problems in science that you could tackle with a 1,000 logical cubit machine? So with a 1,000 logical qubit machine, we certainly can do many, useful things. So we can start to then, simulate processes relevant to drug development.
We develop algorithms that would help, with the design of nuclear fusion reactors. There would be applications in making batteries for electric cars better in the sense of you can charge them quicker or they are lighter, less dangerous to burn. I mean, wouldn't it be awesome if you could have airplanes as that operate on ion air batteries? In principle, that's possible. Those designs have a higher energy density like kerosene. But today, they're still too brittle.
You can't reliably put those into airplanes yet. But with, a quantum computer, you can hasten the development of such a device like, batteries for airplanes because you don't today, if there's an electrochemistry engineer and she has an idea, oh, I think this is a better cathode material. Let's put this in. The only way to test it today is to build this battery, take it to her to lab, and measure it. That's a very slow process.
With a quantum computer, you can simulate this in silico, so to speak, where you now, simulate, the properties of this battery and see how fast it would charge or how quickly, electrons diffuse. And then only the very best designs, that checked out, you take into the lab, measure in practice how well this would work. I see. And and what about, I suppose, applications that maybe people would associate with Google?
Things like, information processing, the optimization of of searches, and, I suppose AI when it comes to dealing with large quantities of information. Could a a 1,000, logical cubit machine be useful for that? Or would you really need lots more cubits, logical cubits to to do those sort of applications? Yeah. So what we talked about so far is this, application area is referred to as quantum simulation. And we often refer to this as Richard Feynman's, killer app.
Because it was Feynman who famously, realized for the first time that it's actually his quote. Nature is not classical, and damn it. And if you wanna simulate nature, you better make it quantum mechanical. So that's questions like the dynamics of, chemical reactions or properties of magnetic materials or making very low resistance, materials for electronics.
And all these problems entail quantum phenomena or these are systems where quantum effects are important and simulating those well is sort of the baseline killer app for quantum computers. But you're in quite right. It's not limited to this at all. Today, we know about 60, algorithms that have a scaling advantage, which means, as, the problems get larger, the quantum computer
can do it more efficiently. Meaning, it can do the quantum algorithms, can do it with fewer, sometimes way fewer steps than a classical computer. And way fewer steps, I mean, can be exponential reduction or a quadratic, reduction in the number of steps you need. And, for example, with optimization, which is another killer app because optimization problems are so pervasive. They are, key in machine learning. They're key in engineering.
They're important in finance. There's hardly any area that doesn't, require the solution of optimization problems. And we have known since the nineties that, for any optimization problem, you at least get a quadratic, speed up in scaling. But quadratic is not as good as exponential and therefore, it would put, quantum enhanced optimization rather far out. But to our delight, our team has developed a new algorithm that's called the DQI, algorithms, stands for decoded quantum interference.
And this algorithm, seems to give us an exponential speed up in optimization for certain classes of optimization problems. We don't quite understand yet which classes those are, but if this were to pan out, this would be super exciting because you can think of optimization as puzzle solving. And let's say if you wanted to build an AI, then, of course, an AI that is better in puzzle solving, that is the one you
will wanna have. So, therefore, my prediction is that in the future, if you have a quantum AI playing chess or go against an AI, the quantum AI will win. I see. And and because of that exponential, effect, do do does that mean that you could conceivably implement those algorithms on a a 1,000 cubit machine? You you might not have to wait until you've got 10,000 logical cubits or a 1000000 logical cubits. That the it's so efficient that you could implement it on a small quantum computer.
We should definitely see first compelling examples of optimization problems in the range of, 1,000 variables. Of course, more is better if you could have problems with, 10,000 or, 100,000 variables. That would be better. But but certain optimization problems are very hard, and you definitely can find the optimal solutions for a 1,000 variable problem. And if we could show that the quantum computer can find better solutions, don't wanna mislead, the audience here.
It is not known today, and many suspect it is actually not correct, that quantum computers can solve very hard optimization problems perfectly either. But what they often can do is they give you a much better approximate solutions than what is classically attainable. So they solve the puzzles not necessarily to full optimality, but they can solve the puzzles better than classical computers. I see. And I don't think, Hartmut, I I haven't asked you, specifically about
the Willow chip. So how many, physical cubits, does it integrate? So the Willow chip has a 105, physical cubits, and they have, very high quality. The physical error rates, let's say, for our 2 cubit gate gates is just 1 in a 1000. So this, allows to run the most complex quantum algorithms today on the Willow chip. Well, that's great, Hartmut. Thanks so much for coming on the podcast. Oh, most welcome. Pleasure to be on your show.
You're listening to 1 of 2 podcasts with our breakthrough of the year winners. The other features Harvard University's Mikhail Lukin and Dolev Blufstein, and you can find it on the Physics World website, or at your favorite podcast provider. You can also read more about our top 10 breakthroughs of 2024 on the Physics World website. This served as the shortlist for our breakthrough of the year award, and it covers a range of fantastic research in Physics. So do check it out.
I'm afraid that's all the time we have for this week's podcast. Thanks to Hartmut Nevin for joining me today, and a special thanks to our producer Fred Ailes. Physics World's coverage of the breakthrough of the year is supported by Reports on Progress in Physics, which offers unparalleled visibility for your groundbreaking research. You can find the journal at iopscience.org.