#329 Izhar Medalsy: How AI Solves Quantum Computing's Biggest Problem

0:00The idea is that if classical devices could mimic quantum devices, then we wouldn't need quantum devices. So going back to the digital twin, those qubits, you're right, they're notoriously noisy. They're very unstable. That's the beauty of it, because that's kind of the core of what quantum mechanics is. That you can get to 80 % accuracy with Shores on factoring the number 21. Is that right? We took this knowledge that we gained from our digital twin and implemented it on the IBM platform and got to 99 % accuracy on Shor's algorithm. Can you start by introducing yourself and give your background insofar as it's relevant, where you studied, what you studied, and how you came to quantum elements?

0:48And then we'll talk about quantum elements use of large scale digital twins and AI to make noisy real world quantum hardware useful. Absolutely. Thanks for having me, Craig. I studied physics and chemistry back in Israel and focusing on nanoscale computational devices using quantum dots and other components. And these were kind of the early days of, you know, those kind of nanoscale computational devices. I did my postdoc in Switzerland at ETH Zurich, kind of diving deeper into the field. Started my career in the industry as a research scientist and very quickly switched to a business dev and product and took the product that actually developed as a scientist to the market.

1:51And that was quite a thrilling and interesting adventure and switched to a couple other companies and built my first startup for a few years. It was a very interesting journey, but maybe for a different podcast. And a few years back, alongside my two amazing co-founders, Professor Daniel Lidar from USC, who is really one of the world leaders in quantum error correction and quantum algorithms, and Professor Amir Yacobi from Harvard, who is a leading experimental scientist in quantum, we realized that in order to deal with those very unstable systems, meaning those quantum devices, you need to be able to use the ability to augment those systems using classical devices, meaning creating a digital twins of those systems and best in breed classical tools that are used today for software development in order to control them, accelerate their development and ultimately enable end user applications on quantum.

3:06Yeah. And when we spoke before, I was asking whether this is, when you say using classical hardware, whether this is using classical computers to run quantum algorithms, which there's a lot of work in that field. But from my understanding, you're building digital twins of quantum hardware and then using AI to optimize that hardware. And you have some pretty interesting results. Is that right? And if so, can you explain that? That's absolutely right. The idea is that if classical devices could mimic quantum devices, then we wouldn't need quantum devices. On the other hand, in order to allow AI and other software components to accelerate the development of quantum, you need to give those developers a platform that mimics those quantum systems.

4:18right as we all know quantum systems are not an extension of classical devices they're completely different species and completely different technology so what we've done is we really went back to first principles and solved those equations that describe those quantum systems We solve those equations on large classical computers, you know, supercomputers. So now you have a very realistic representation of any given quantum hardware. And it allows you to experiment and solve and develop different layers of the software stack that are specific to the hardware that you're interested in. Meaning if you are sitting in a lab developing your own system, you can use this digital representation to solve problems on how the system works and how to optimize it, etc.

5:18On the other hand, if you're sitting in a company that would like to develop an application, which hardware will you choose? It's very hard. They're very expensive. They're very scarce. So if we're giving you a realistic representation of the system, you now have the freedom to experiment, develop, and augment different systems and understand how to work on them and ultimately execute your application with certainty and confidence. And when you say build a digital twin, can you describe that as is this pure math or is it something more tangible? So there are two components to building a digital twin.

6:04There is the physics part and there is the algorithmic part. The first thing is looking at the system in front of you and being able to describe the math that is behind its working principles. Specifically in quantum systems, you need to be able to describe what is the modality in front of you. Are you using superconductors or maybe you're using ions, right? There are many different ways today of realizing quantum devices. Just to put it in perspective, there are about 70 different quantum hardware companies out there, each one of them with their own kind of unique flavor. So the first thing is to look at the system that you're interested in creating this digital twin for and saying, OK, what are the governing principles of those systems?

6:56What are the equations that if I solve them will give me a reliable prediction of how this system behaves? and in the same way that you look at any physical system think about the spring right in order to understand how a spring works you need to be able to solve the spring equation of this specific spring that you're interested in so the same principle now the other thing is okay you're now able to solve those equations but they're very expensive computationally so how do you do it efficiently in a way that you can enlarge them so once again to put it in perspective There are some tools out there, scientific tools, that can solve one or two qubits of those digital twins.

7:42But what can you do when you're able now to solve 50 or 100 of those qubits? You can now tap into every problem that the industry is tackling and provide the right tools because you solve the system at scale to a point where it's now very meaningful. So these are those two components, the physical representation, the ability to scale it reliably. And when you talk about solving for individual qubits, I mean, quantum is notoriously hard to understand. Qubits are hard to understand. And are you talking about, does the digital twin simulate individual qubits? Or are you talking about optimizing all of the things around, you know, for sampling or for temperature control or for all the other, you know, isolating or eliminating noise?

8:54all the other things happening around the qubit? Or are you actually doing something with the qubit? That's a fantastic question. The third piece that comes into play when you need to provide users with a digital twin is the ability to also take into account the environment around the system, right? What is the noise that is affecting the system? and how does it evolve over time as you are operating this machine? So let's take an analogy from a different world and then come back to quantum. Let's say you want to develop your next generation airplane. The kernel that you're using in order to understand how the air is flowing over the wing of this new generation of airplane that you want to design is solving flow dynamics, right?

9:53You want to understand how those particles are moving across the wing. So you start with a single particle, you kind of solve the equation. And now the secret sauce is how do you enlarge it to represent millions of those particles that are flowing over the wing? And if you want to take it to the next level, what happens to the flow where the airplane is stalling or maybe when there is some non-laminar flow or some issues with the weather, you know, kind of you have some other phenomenas that are environmental now that are affecting this airplane. So going back to the digital twin, those qubits, you're right, they're notoriously noisy.

10:40They're very unstable. That's the beauty of it because that's kind of the core of what quantum mechanics is it's a statistical system but it's also very hard because we need them to be quiet enough in order to do computation so what we're doing at quantum elements we're solving the equations that not only governs how those qubits are working but also what are the environmental conditions that are prohibiting them from working correctly with low enough noise and And we give the user this kind of recipe or all of those components that are now jeopardizing the performance of their qubits. Now that they have at their disposal all of those different knobs that control the performance of their qubits, they can go ahead and develop the mitigation in order to make them work correctly.

11:38That's exactly the word that the industry is using. Air mitigation, air correction, air suppression. These are all tools that are built on top of the engine that we're bringing. And because we're able to do it with such high fidelity and accuracy, and it's such a large scale, it now uses that the users are benefits. Benefit is that they can now do it at scale with very high certainty. And this is tuned to each particular hardware or the modality, as you put it, that the hardware is using, whether it's superconducting qubits or trapped ion qubits. So what data do you need from the hardware provider to build the digital twin?

12:38And it seems to me it would be not only the hardware, but where the hardware is positioned in what kind of a room and what kind of a building and all of that. This is exactly it. And it really speaks, your question really speaks to the crux of what we're doing. So there are two aspects here. As you go up in the software stack, if we take the classical analogy, if you're developing an app, you need kind of some sort of a translational layer to the hardware that is underneath it, right? It can be very up in the stack. It can be very abstract. As you go deeper, you need to know more and more about the hardware to a point that if you're developing an operating system for, let's say, a GPU or a CPU, you really need to understand all the intricacies and architecture of your system.

13:35So same with quantum. When we are working with hardware manufacturers, we would usually have a very intimate relationship with them. We will work with them very closely to refine this model so it will truly represent everything that is going on with their system because it is used by their own developer in order to refine maybe the architecture or the design or how those qubits are working from an operational point of view. um but when we're working with customers who are upper in the stack that are maybe interested in application the models can be more abstracted and in that regards we don't need much from the hardware manufacturers um if we take another you know analogy outside of quantum let's say you're developing um formula one simulator you have two paths as the guy who's designing the simulator.

14:40You can go to McLaren and sit with engineers and ask for all the design files and everything. And they would say, okay, we need to sign an NDA and we need to have some relationship. But if you accomplish that, your model is not really true to life. And probably the people using it will be the Formula One drivers of McLaren, right? But there is another path. you can borrow a McLaren, run a couple of laps on the track and kind of record the acceleration and the drift and everything, and use that to have a more maybe abstract model of your car that will be used by novice drivers, or maybe people are interested in developing a self-driving car, not necessarily the next generation of McLaren.

15:27So you can see that you can tap into different users personas and user different users and participants in the quantum workforce based on what their needs are and how accurate this twin needs to be so yeah so i mean you're you answered this question but maybe in in more layman's language how how do you validate that your twin is a true representation of say IBM heavy hex device or something like that. So let me give you an example. I think one of the most famous algorithms out there is Shor's algorithm. It's an algorithm that allows you to factor a number. So what we've done is we went to the IBM platform as users and kind of looked at the published information that they have on their website on the performance of their system.

16:27We know what superconducting qubits are. We understand what the models are. So we wanted to see the relationship between how those systems are behaving versus how Shor's algorithm is behaving in terms of its accuracy, meaning when you're executing Shor's algorithm, how accurate is the result? So we went and executed Shor's algorithm on IBM's system and got to about 80 % accuracy. And we've suspected that there is something in the relationship between those qubits. and we know how to solve this relationship. It's called crosstalk, meaning when those qubits are influencing each other, right? In an ideal world, every operation that you do on one qubit should not influence the other.

17:23So what we've done is we've simulated thousands of times those IBM qubits with different crosstalk combination because if you had to do it on hardware, you will have to manufacture a new QPU with new crosstalk, right? And then we realized that there is this sweet spot where we can inject a process called air suppression that can give us the best performance on Shoros algorithm for the specific IBM system. We took this knowledge that we gained from our digital twin and implemented it on the IBM platform and got to 99 % accuracy on Shores algorithm. So if the highest reported accuracy on Shores in literature is about 70%, when we submitted the circuit as is on IBM, we got 80%.

18:17When we implemented based on the knowledge we gained from our digital twin, the remedy for this crosstalk, we got to 99 % accuracy. So you see that the ability to close loop, to validate and to get the benefit of using this technology can be very quick. On the other hand, when we're working with a hardware manufacturer and we're working with one of the large public companies very intimately, we're literally sitting with their scientists and gauging how accurate is our simulation and digital twin against their experimental result and refining the model based uh yeah that's that's fascinating uh how how sensitive is that 99 percent to drift in the underlying hardware parameters i mean if ibm recalculates its chip tomorrow do you have to relearn everything or or are these uh these optimizations applicable regardless there are there are a few layers here um when you look at at qubits um your main at performance you're mainly looking at as what is known as decoherence how quickly are those qubits moving away from their optimal condition yeah um these numbers are publishable and they would usually change based on the calibration process that each of those manufacturers is doing, usually on a daily basis.

20:03There are other more kind of intimate parameters that will govern the performance of those qubits, which we can extract just by using the system as any other user. So to your point, the model is the same model, but the noise and performance parameters have to be updated daily in order to keep track of how the system evolves day over day. And that's the benefit of using this kind of very intricate physical understanding of the system is that once you solved it from first principle, the only thing left is to optimize the daily parameters that change due to environmental conditions you know things drift and temperatures we need to remember those superconducting qubits are sitting in temperatures that are colder than space right so it requires just a milli kelvin degree change where everything starts to drift and change so yeah that's how we do it yeah yeah and so what's uh which noise channels and hardware parameters Are you modeling?

21:13I mean, or is this all approximation? So these are models. These are physical models. They're known as master equations. They're really solving the quantum system from first principles, right, from describing it from the physics side of things. And there are a few noise models that are very important. The first one is known as T1, T2, and T-star. These are looking more at the rate of noise accumulation or decoherence in those systems. We can do that. There is 1 over F noise. That's another one. uh there is crosstalk and crosstalk is a very notorious one uh think about this game of whack a mole right uh you're dealing with one guy and you think you're done you calibrated this qubit but then you realize that in the process of calibrating this one the other guy went out of calibration because they're influencing each other so we know how to model to model that and And it's a very important one because when you have a few qubits, then, okay, you tap this one and you kind of think about the knobs of, you know, this analog machines.

22:38You're turning them and you kind of, you have the secret sauce, you're the expert. But when you have a hundred, it becomes a bit harder. When you have a thousand, we'll forget it. So it probably would lead us to the next topic, which is how is AI kind of tapping into all of them. Yeah, that's right. And the last one is leakage. Leakage is also a very interesting one because you have to remember, you know, we're going back to a bit to quantum physics. Most of those qubits are using what is known as a two-level system. You kind of move between, let's say, let's call it a zero and a one. It's a bit more than that, but a zero and a one.

23:14But sometimes there is a three and there is a four, and, you know, those things are unstable. So that's another thing we can model. Yeah. And for listeners who are not quantum experts, and for me, who's not a quantum expert. So you showed that you can get to 80 % accuracy with Shores on factoring the number 21. Is that right? Yeah, so we can take it from 80 to 99%. 99, I'm sorry, to 99. Yeah, 80 was where the industry was at.

24:04Can you, yeah, explain what factoring 21 means. So I know it sounds like a very small number, right? My nine-year-old can probably factor 21. but it's more of a way to demonstrate that, or to illustrate the fact that quantum can do something very unique. It can solve those factorization problems that ultimately tap into encryption in a very efficient way computationally. So the reason we are all relying on RSA encryption is because if the bit string is big enough for classical devices, it will take the age of the universe to factor it and by doing that to break the encryption. Quantum computers are working in, you can think about it as a parallel way.

25:06It's almost like we lay in front of you all of those possibilities and the system will converge to the accurate solution in one go. And the way we've, to demonstrate that is, you know, Shor's algorithm is kind of solving this problem in an effective way. And that's kind of gave the industry or this field the acceleration and the confidence that we can now use quantum computers to deal with real-world problems. The problem being still, and that takes us to the state of the industry today, is that to develop those algorithms at scale, you need very unique capabilities and approaches. So we're still at very early stages, meaning those numbers are still small, But I'm pretty confident that as the hardware evolves and the number of qubits grow, we will start seeing companies and scientists that are solving shores or factoring shores on much larger numbers, getting us closer to the point where quantum can really start tapping into the encryption industry and disrupting it in the way that we think it will.

26:36Yeah. Okay. And how is AI applied? I understand the digital twin concept, but then how are you using AI? Yeah. And are they standard models or are you building bespoke models? Are they generative models? So I want to talk about the here and now first, and then we can maybe think about the next. One of the things that AI needs, and I'm sure your audience can appreciate that, is large-scale curated data, labeled data that is actionable and can be used to solve a specific problem. One of the things that we can agree on after a 30-minute discussion is that quantum is scarce, expensive, and exotic.

27:33meaning that each modality has its own intricacies. They're not easy to come by. And even if you go ahead and dedicate a machine to train a specific AI model, the next day you can wake up and this generation of machines is already history and you need to repeat the process again. The advantage that you have with the digital twin is that you're not bound by any physical hardware. You can not only create a representation of the current state, but also of future state to a system scale that is big enough that allows you to generate massive amount of data to be able to address some of the issues that the industry is facing today.

28:26the most present and critical issue that we're dealing with is noise right and errors so one of the things that we are doing and we're gonna release it very soon with our partner is the ability to do large-scale error correction using a digital twin now Now, if you want to develop, let's say you're a scientist in a company that is really interested in now dealing with errors on a large scale, and you want to develop a decoder that is solving those errors, not only for the current generation of your machines, but also for the future one, You need to be able to generate a lot of labeled data that will allow you to train this AI-based system and decoder in a robust way.

29:24So we're giving you the platform to do it very economically and in a very robust way. One of the things that I didn't mention about using those large-scale simulator versus hardware is people sometimes hear this notion of shots in a quantum computer. You need to run multiple shots. Sometimes it's 10 ,000. Sometimes it's more. Why? Because when you run a quantum computational cycle, you get a snapshot into the statistical representation of the accurate answer, meaning you need to run enough cycles in order for the answer ultimately to be statistically significant enough to represent the right answer.

30:14That's how kind of wild quantum is. When you're running our digital twins, one experiment gives you the full picture of all the statistical representation of your answer. It's called the density matrix. It gives you the whole universe of solutions. So when you have this power and at scale with the right architectural configuration of errors, you can now, as the end user, say, okay, you know what? I have a new idea. I think ML, AI, reinforcement learning, whatever. if I have enough labeled data that is representative of the system that I'm interested in, I can now build the engine that will allow me to solve those future problems.

31:03So think about us as this development platform that gives you the flexibility to generate and augment and use large-scale data for your AI needs. And that's one of the things that excites me the most. And on the labeling, what exactly are you labeling? What's an example of the labels that you're applying and how do you curate them? So think about, you know, if you know that you need to do those, let's say, error correction cycles. In error correction, you're doing a process that is known as syndrome extraction. You want to understand what are the syndromes that are governing your errors and how to deal with them.

31:53If you as the guide designing the system can change the knobs of your hardware virtually, see how it taps into the output of the error. You now have a direct link between hardware vulnerability and error accumulation. So now you've connected the dots, right? So if your cost function is, I want to minimize the number of errors, and you're now curating the data in a sophisticated enough way that allows you to start connecting the dots between the governing principles of what generates the error and how the errors are manifested on this large scale, you now have the ability to start minimizing the governing components that ultimately are jeopardizing the performance of your system.

32:49So because, you know, one of the things that we always say about the quantum system is the moment you ask the quantum system a question, it collapses, right? it kind of the the notorious thing about quantum is you cannot stop at a given point and say okay what is your state if you're asking the quantum system what is your state you've you've killed the process the process will not continue you cannot say okay what's your state and then okay you can now continue but you can do that with a simulator so it gives you this superpower to stop at every given point, understand the underlying mechanisms, and continue.

33:29And this allows you to be very selective on the parameter space that you are using, how you're training, how you're labeling. You almost have the full recipe of what is going on, with the caveat that obviously you need to make sure that then you go and test it on hardware, and et cetera, but at least you have the control knobs at your disposal. Yeah. And can you sort of walk us through the pipeline from a user uploading an algorithm through the simulation engine to the AI training on those runs and proposing a remedy for the hardware and optimization for the hardware? So if you join our platform today, you will be asked a few questions.

34:27We will ask you, you know, what kind of experiment are you running? So let's, for an example, move away from error correction, and let's talk for a second about error suppression. We'll describe what error suppression is in a minute. So what kind of experiment are you running? Are you connecting to a physical machine or do you want to create your own virtual machine? If you're connected to a physical machine, we can extract the system parameters for you and kind of feed them into our model. What kind of modality are you using? Let's focus for a moment on superconducting qubits. And what is your system size?

35:08Let's assume you want to use a grid of five by five. Now we will give you a canvas. You will have a canvas, a digital canvas with those qubits. You can now control the connectivity between those qubits, and you can literally go to each qubit and update the system parameters, the frequency, the T1, the T2, the leakage, the crosstalk, all of those things that I talked about. And from that point on, this system will behave as a physical system. You can upload a circuit. So there is a tab. You press the tab. You upload your quantum circuit. And now you can run this simulation. And we allow you to inject pulses into your circuit to deal with the different noise sources that you have.

35:56So now think about the following process. We give you the tools to create your machine. We give you the canvas to run your circuit in, and it's a very fine detail, but it's important. The circuit runs in the time evolution domain. It means it behaves in the same way it does on hardware. I'm saying that because all of the simulators out there work as discrete systems. They're not tied to a physical entity. So now you can run this circuit, and you can inject different remedies. So now let's think about a model that you can develop. You can think about the iteration and the placement of those injection pulses and now start running multiple experiments, thousands of iterations where you're optimizing location, shapes, and components, where all of the time you're monitoring the fidelity of your circuit, the performance of the circuit.

37:03And when you're hitting the optimal point, you know now that you've kind of closed the loop on this cycle. From that point on, you can now start thinking about training a model that will find the next iteration when conditions changes to hit the same performance level and again and again and again. So it gives you the flexibility. It gives you the canvas to do all of this work. And obviously, there is a huge cost benefit here, right? When you're working on real quantum hardware today, if you go to the leading providers, you would probably pay a few thousands of dollars an hour to run a circuit.

37:44You can do it in a much more economic way on our platform. And when you're gaining your confidence, then obviously you go to the hardware, but you understand what you're doing and you understand the root cause for the things that work and do not work. And can you give sort of a before? I mean, you did earlier in the first time we spoke, I mean, we did, for example, the 80 % to 99 % on factoring 21 with Shor's algorithm. But can you give some other before and after results? Yeah. Absolutely. One of the things that the industry is moving towards is logical qubits. It means when can we take those physical systems?