Material Science with Houlong Zhuang at Q2B Paris

The New Quantum Era, a podcast by Sebastian Hassinger. And Kevin Rowney.

Hello, and welcome, listeners, to a special episode of the podcast. It's just me, Sebastian. Kevin is unavailable this week, and we've been trying to keep to a more regular biweekly schedule, so we agreed that I should just go ahead. I was in Paris last week for q2b 2024. It was the 2nd year of the conference.

So let me introduce myself first. Yeah. So my name is Hongmong Zhuang. I'm assistant professor from Arizona State University. So how did I get into quantum computing? I started as, when I first joined ASU, I started teaching a very basic mathematic class. I told you yesterday, it's linear algebra. And my initial purpose, teaching that class was I was very interested in machine learning. Right. And machine learning basically is applied

Right.

Linear algebra. So yeah. I yeah. I had a chance to teach that class. Right. And then after I taught that class, I realized, oh, linear algebra can be applied to so many different

It's kind of like the native language of cubits. Right? I mean, that's kind of the you know, instead of Boolean algebra, you're actually using linear algebra when you're when you're doing a quantum algorithm.

Exactly. So for example, all the gates I mean, people call gates in quantum computing. Right. But, essentially, they are unitary unitary matrices in linear algebra.

Right.

So if you know linear algebra very well, then it's automatically you can

It's linear algebra machine.

Yeah. And then I was trained in density functional simulation, computational material design. Right. Right. Basically, you apply linear algebra again Right. To solve this eigenvalue problem.

Right.

When you solve this eigenvalue problem, you know that I, you know the eigenvalues. And eigenvalues can be interpreted as the energy network of electrons.

Right.

Then even you don't do the experiment, then you can predict the properties of materials.

Right. Right. Yeah. When when did the DFT sort of approach to material science? When was that, you know, invented? So or explore or discovered, I guess, science.

It was it it it dates back to 1960. The original paper, if I remember correctly, they were published in 19 sixties. Mhmm. And 2, one of the inventors so he received Nobel Prize in chemistry in 19 eighties. Right. Right. Yeah. And this is a very, very efficient way to solve this context running or And

but it is like, it's can be very difficult computationally, though. Right? Like, on a classical computer, a DFT problem can very easily outscale the the size of the computer you have, which is why quantum computing is sort of interesting for material science.

Yeah. It depends on, like, the level of theory. Mhmm. Right? Because DFT standard DFT so one of the most challenging power we call is training correlation function. Essentially, it's how electron interact with each other. Mhmm. Right? They interact each other in 2 ways. One is the classical way. Mhmm. So classic way by classical way, I I mean, you are negative charge. I'm also negative charge. So we apply Yeah. Now.

Right.

So then you keep on adding more and more approximation. So it starts to deviate from original.

Yeah. It's similar to the problem in chemistry. Like, computational chemistry, the you're building approximation on top of approximation. Eventually, your your accuracy your results are just not gonna have to be useful. So it's similar with material science.

Yeah. Yeah.

Yeah. So And if you have an accurate DFT calculation, what does that tell you about the what does that predict about the materials characteristics?

Oh, that's very good question. So, essentially, all you need for input to, compute classical computer simulation is the atomic arrangements. Mhmm. Right? So you can, for example, you can input the, atomic arrangement on carbon actin Mhmm. In the way of diamond. Mhmm. Mhmm. Right? And you can also input the arrangement of carbon actin in a way, graphite.

Mhmm. Right.

Right? Right. And then you can yeah, you can input those 2 structures to DMT program, and DMT program will give you 2 energies.

I see.

And then it will tell you which energy is, no energy.

Right. Right.

So in principle, no energy means more stable structure.

Right. Okay.

Yeah. So which means that you can input the arbitrary combination of atoms Yeah. Or species.

Yeah. Yeah.

Yeah. It will give you some energy. Yeah.

And from that energy, can you can derive sort of the characters. Like, diamond is a very hard substance and and graphite is softer. Like, that that can be derived from the the the the the from those energy values?

Yeah. Actually, so there are several mainstream DLD programs for a solid state camp community. Mhmm. For example, Conde Espresso, VAS. Right. Yeah. So or from Cambridge.

Okay.

Actually, all those programs, they initially they are called the total energy based.

Mhmm. Okay.

We would you go back to your question? Okay. So everything on many, many properties can be derived from

total energy. That's cool.

So, for example, elastic constants can be written in terms of some derivatives of total energy.

That's cool. That's really cool. So okay. So Yeah. So that's a a foundation on DFT material science. So you're now, looking at a way of using quantum computers to, to solve certain aspects of or challenges around DFT. Right?

Yes.

And how how are

you doing that?

So We're hoping to do that.

Yeah. Hoping to do that. Yeah. I I mentioned earlier, the most difficult part is the quantum interaction one molecule interact chemically interact with some substrate material Yeah. You know, some charge transfer.

Right.

And you cannot use this standard version of this functional to describe these interactions. Okay. So in that sense, you need to use this very complicated many body theory.

What's an example of that that type of molecule interacting with the substrate? What is that like a like a composite material or the with the layers? Is that that sort of thing? Oh, oh, many examples.

So for example, use carbon dioxide molecule is

Okay.

Chemically of the salt. So I see. Ionic liquid. Yeah. When they as far as an age start to have some charge transfer, your electron become mine. My initial become yours.

I see.

Right? So in that's wrong in that's wrong, in interaction start to have to be very complicated.

Right. Okay.

Yes.

Interesting. So okay. So it's broadly applicable. It's very you have to do the quantum calculation, which is quite complicated. So you're looking for a computationally efficient way to do that on on quantum computers.

And then you can of course, you can do, very complicated quantum chemistry by adding more and more many body. We call it many body facts. Right. And and quantum computers, it they are promising to help solve this, to implement this, for example, CCSP is couple cluster method. Okay.

Yeah. Interesting. Interesting. So okay. So what's what is unique about your approach then? What's what's your research focused on?

Oh, so my talk yesterday was about a system called a high entropy material. Mhmm. So it's different from conventional alloy. So conventional alloy, like in a steel Mhmm. Yeah. Make, which maintenance chair you are sitting. Right.

Yeah. Yeah. I'm familiar with steel.

Yeah. Yeah. It's dominant atomized, iron, maybe 90%. I see. And there's some tiny amount

of current.

I see. Okay. Zero less than 0.3%. Okay. We call it conventional alloy. Okay. One dominant. Right.

Okay. And then

very small amount of elements. Got it. And in contrast to this conventional element, we have this high end of the alloy, so which means that you have many, many elements. So the original experiment doing this, high entropy material system is one undergraduate student from UK University. So he was doing some undergraduate thesis. So his adviser, asked him to mix more than 20 elements together. Just mix together to see what you can get out of that.

It reminds me of, you know, sort of being a kid and, like, mixing stuff in the kitchen to make it like a magic potion. Just throwing a bunch of spices in and seeing what happens.

Exactly. Like, the mall is different. Right? Yeah. Just now you you go to some buffet. You eat. So many varieties that you mix together.

Make a new food.

Yeah. Or may maybe you can make your stomach upset.

Okay. So what what did he get when he mixed together the 20 elements?

The original article actually not much. Yeah. Like, in the waiting Yeah. It just very yeah. So exactly. Just, yeah, just a new way. But after maybe maybe on a in the same year, a Taiwanese group, may independently did similar experiment, a mix different element. Mhmm. Yeah. And then from then on, more and more people start with all these so called computational space. Mhmm. But you can imagine each element is one axis in a high dimensional space.

Right. Right. Which starts to sound like the Hilbert space.

Exactly. Exactly. Yeah. Yeah. It's yeah. Here, but it's very exact.

As soon as you start saying high dimensionality, I think of of, you know, cubits all in entanglement.

Yeah. So you think hydrogen is x axis. Right? Helium is y axis. Right. Right? In the same group of elements. And they may they may not necessarily have orthogonal axis, but they may be closed. Right.

Yeah. Very complicated.

Yeah. And then now this high end of the material, this family of material become very popular in material science community. You can find almost high and beyond everything. High and beyond, superconductor. High and beyond, to be semiconductor. Interesting. And to be catalysis. Interesting.

And what what why is it becoming so popular? Because there's so much that's unknown to be to be discovered about it. Is that

Yeah. You can imagine this each each anime is, like a is a pixel. Right?

Okay. Yeah.

And then you can try to paint this a new material Cool. Using mixture of That's

really cool.

Things always. But, I

mean, it sounds like I mean, you said yesterday, in the talk that there's a combinatorial problem. Right? Yeah. You got n by n or m by n or whatever. There's, like, multi multi current kind of combinations just to start with. So is that you're starting with trying to narrow down the possibilities that are are for your sort of research space or your experimental space. Is that right?

Yes. So we originally start from classical machine learning. Mhmm. So we have some, training data set from experiment saying, oh, which combination of elements can give you some certain properties. Right. And then we can train the deep learning model. Right. And then so that next time you have a new mixture of elements Right. I can I can predict Yeah? Yeah. For you in a new property. Yeah. Yeah. And we started kind of very early. So we we I had a very good masters in the marketing business.

Yeah. That's pretty good for 4 years. Exactly. That's great. Yeah. That's great. So that was the classical machine learning. Mhmm. Are you it was the next step to start adding in quantum simulation, or were you applying quantum to potentially get a better machine learning model?

Yeah. We had, we have already started the quantum machine learning model trying to achieve the same accuracy using available quantum simulator Right. Right. And quantum hardware. Actually, yesterday, I present some results. Right. But we used a very adding very rudimentary encoding of the Mhmm. Entropy analytica. I So we simply we only use the chemical formulas.

I see.

Yeah. But I think we can do a better job. For example, give some more input or more some more chemical intuition. Mhmm. Yeah. So in our world, we haven't considered, oh, the relation if on a if some animal in the same group, right, they shouldn't be orthogonal in each other. Right? But the in the you know, what we are now, we we assume is Okay. Okay. Orthogonal in each other.

So better better sort of, whatever, starting point, starting assumptions

about Yeah. Yeah. Or different ways of encoding this. Interesting. Yeah.

And Is that challenging, by the way, with quantum machine learning. I mean, the the the preparation the state preparation is often really a lot of the overhead. Right? Because it's you have to fit that high dimensionality into a very small number of cubits. Is that is that sort of part of the challenge?

Yeah. It's part of the challenge. But I think another challenge comes from the noise of the

Mhmm. Of course.

Yeah. And each material have, must can must be able to be retained in terms of its competition or chemical formula. Right? Right. We only we we have about 100 elements. Mhmm. Which means that if we do 2 to the power of 7, you can you can cover all the elements in the

Oh, to the power of 7 cubits? Or

Yeah. 2 7 cubits.

Oh, 7. Wow.

Yeah. 7 cubits. Because you can use, one string of 0 Right. Numbers and 1

It's all vectors.

Right? Yeah. Yeah. Yeah. I mean, the algebra. Right. Right. Yeah.

That's so I mean, I find that so fascinating. I think coming from a classical computing background, that's the hardest thing to sort of wrap your head around is is thinking in vectors, thinking in encoded vectors. And, of course, as you said, that's happening in classical in the area of machine learning, with with linear algebra. But it's still, like algorithmically, that feels like the the biggest barrier to entry for for people with a classical background is to start thinking in those that matrix, multiplication kind of space.

Yes.

Right? I mean, like, Peter Shor's big, you know, sort of breakthrough is is having that incredible, you know, quantum phase estimation kind of black box where the the the, you know, the factorization happens almost magically. It's Exactly. Because of the way you approach it. It's such it's such a cool part of the whole theory.

Yeah. I taught a class and I assigned Peter Saw's algorithm as one of the humble problems. And one student in my class, so he did the exercise. So he basically, he decomposed 15 into 3 and 5. And he solved the problem, and he wrote me an email saying that he found a very powerful algorithm solving that problem. That's the old that's the first this kind of comment I received from students. That's cool.

Yeah. That's also by the way, that's, IBM Research in 2001, I think. That was the first successfully used of Shor's algorithm on 7 qubits on an NMR device, and they factored 15 to 35.

3 and 5.

Yeah. Which doesn't sound like that impressive a task, but until you realize how they're doing it. Yeah. The algorithm

is pretty Yeah. The algorithm is really neat and meaningful.

It really is.

Private. Wow. This guy deserves a genius.

I know. And then and then coming up with the 1st error correction scheme, like, a year later, that's it was, it was quite a quite a productive year for Peter.

Yeah. He had a he had a track record in this kind of research. Right? Because before he he discovered or invented this algorithm, he was already very famous expert in

Oh, really? Theory.

Okay. So he published a number of papers or numbers.

Interesting. That's funny because my focus is on quantum computing. I don't know anything about Peter Schwer before.

He was already the same as he got.

That's interesting. Yeah. It makes sense.

You probably that

seem brilliant. You don't just be that brilliant overnight. Right? Yeah.

So that's probably why, like, he naturally. Right. Yeah. I cannot believe that.

But anyway, returning your research. Let's get back to the material science. Okay. So you had this classical machine learning, sort of down selector attempt to predict the the the properties. You're now looking at quantum machine learning, replication or or, you know, enhancement of that potentially.

Yeah.

So then you were talking about from that smaller pool of potential materials is the next step to try to simulate using a quantum computer, the the DFT

type of We we are going to first simulate using DFT, but it's also where we can use quantum computer. Okay. Because when you do DLT simulation, you need to have a input,

Okay.

Yeah. Gas, arrange Right. Right. Atoms. Right. Right? And then you have many elements in we call it supercell. Mhmm. Essentially, you can imagine this is a cubic grid. Right? Mhmm. And in each grid point, you decorate with some element. Mhmm. And you there are so many different combinations. You said. Yeah. Yeah. Yeah. Yeah. Yeah. You have different permutation. You're still representing the same compositions. Mhmm. Right? We are not going to do all the permutation.

I see.

To come up with some statistical

Okay. So the machine learning step, whether it's classical or quantum, narrows down the candidates. The quantum steps similar or or sorry, runs predictions of of what that initial decorated state would be, the super cell state?

Yeah. Exactly.

And then and then that's the input to the DFT.

Exactly. I see. Exactly.

Is there a quantum algorithm that's equivalent to DFT? Is there is there a q DFT?

Very good question. I saw some papers. Mhmm. So for example, they apply quantum quantum counterpart of DFT. Quantum, not necessarily DFT. Quantum counterpart, we call it DFT based molecular simulations. Okay. Yeah. Interesting. That's this kind of very general natural extension. Mhmm. Mhmm. Mhmm. Right? We know, quantum gametes, they probably good for something. I can value.

Right. Right.

Right. Right. Yeah.

And then If you wanna simulate a quantum system, you a quantum system.

You need to get the eigenvalue. So Yeah. Yeah.

Interesting. And so, like, what is this possible on on sort of NISC machines, or is this something that would need fault tolerant cubits to an error correction to to scale?

So in principle, you need the

You need fault tolerant?

Yeah. You need the yeah. Like, in the yeah. Like, one slide. Sort of, some, this morning.

Yeah. Right. Yeah. That was, it was QW at Pharmaceutical.

Yeah. Yeah. Yeah. Yeah. So basically, he's doing similar things. Right. But he's doing he's working on this, molecular system.

Right.

So mind disease is mostly solid system, which means you can repeat.

Okay. Interesting. Because the pattern repeat the structure repeats over and over again.

Yeah. I understand. I see. Some of copper. Right? It's Right. Yeah. It's it's periodic.

Right. Right. Interesting. Interesting.

Yeah.

Cool. So okay. So, then what what's sort of the scale of of logical qubits that you would need to do something that's that has clear advantage over any kind of classical approach, do you think?

So so far, I think if I have 7 perfect cubits Oh, really? That's small number. Yeah. So I can I can include everything? Okay.

You were saying? Anybody real Yeah.

Right? As far as they have something Interesting. Quantum circuit. That's really interesting. I think in the paper we published, we saw some preliminary law. So we input the chemical formula, and if it is implemented in quantum simulator, only the bars representing the element showing up. Only other bars. Yeah. And they don't show up. Each bar here means, one string. Okay. 0s and ones. Okay. It's each element. Right.

I see.

Yeah.

So but then so you can simulate 7 perfect cubits. There's no performance advantage on a on a classical simulator, or they're obviously doing just a straight classical way? Or is there is there some is there some, like, algorithm that or approach that you're using that that has advantage in either either class or a quantum way then?

Yeah. So so far, we are looking at the classical Yeah. Content content algorithm. I by classical content algorithm, like, I mean, yeah, like Peter saw Right. Rovers algorithm. Yeah. So we are looking at oh, how can you let's say you have 3 bars showing up by after you do after you many slots. Right. How can you change those 3 bars into 1 bar, and then that one bar representing some kind of encoding? Interesting. Category. Okay. Right? So no matter

It's almost a compression method.

Yeah. And it also all these quadratic

so it's probably Okay. Quadratic search. Yeah. Yeah. Yeah.

That that has been sung by, Grover. Right, Grover. Yeah. Yeah. Yeah. Yeah. I think that's where we can possibly see quantum quantum.

And and what would, like, the ideal sort of outcome in your mind? Is it is it some sort of approach to be able to, you know, efficiently, predict the the alloy composition that would have the high entry entry alloy composition, would have certain characteristics that you you desire, basically, sort of design your materials, if you will. Is that is that kinda what you're trying to do?

Yeah. Exactly. So let's say you train your you train your condensate circuit using, a vital word error or good catalysis or bad catalysis. Mhmm. Mhmm. Right? And then in the end, you will come up with some new common teacher. Right? Kind of like extrapolation. Yeah. I'm very interested in recent, classical machine learning algorithm. DVD Yeah. Yeah. Yeah. It's it's generated models.

Right.

Yeah. You can come up with so you learn the probability distribution of classical data, and then you take one point from the probability distribution, and then you you based on that point, you can come up with some new thing.

I see. And then you would sort of that would be the input into the into the quantum circuit, to to predict the or simulate the those character or the certainly, the the, the supercell.

Yeah. Yeah. And there are some recent research on people have started published some paper on so called quantum attention mechanism.

Mhmm. Mhmm.

Yeah. So, essentially right? So in in classical machine learning, each token representing, for example, each English word Mhmm. Right? So may have some kind of correlation. Mhmm. Mhmm. So for example, you say a sentence. This word

Right.

Pays attention to more attention to another word.

Right. Right. Right.

Right. So you can imagine each word is a particle. Right. Yeah. Or not necessarily a quantum particle. Right. That's what I even think is a classical particle. Right. Right? And then you start to have this, correlation.

That's interesting.

Yeah. Of this sense.

I love I mean, what I really find fascinating is that that you're combining sort of cutting edge classical techniques with these emerging quantum techniques.

I mean,

I feel like that that's going to be you know, when we get to some kind of working quantum advantage for some, you know, material science or chemistry or other scientific physical science application, it is gonna end up being some very tightly bound combination of classical and quantum that actually delivers that that resolve. Right? Exactly. Yeah. Very interesting.

Oh, yeah. So, so the whole overall overall program is called the assessment.

Mhmm.

Yeah. A stands for Africa. And so, basically, this program is trying to collaborate with physicists in Africa That's right. To help promote DFT or electronic structure. Right. Yeah. So it was originally started by, Richard Martin. Mhmm. I think he he retired Mhmm. Yeah, from UIUC. He's very he has a very classic, book, in that Sonic segment.

Right. The name is familiar.

Yeah. He is. Actually, you also mentioned the innovative, ABS innovative fund. Mhmm. Yesterday. Yeah. I think he got a one of the Oh, really? Innovative fund for this electronic tractor. Maybe in the first yeah. Oh, wow. Yeah. In the first version of this innovative

Oh, that's great.

Yeah. So Right there.

So he started this program, sort of collaborating with African physicists. And and what you've done in now 3 years, I think.

So he has been doing this.

Oh, but you you've just taken

you started to be Yeah. I was planning to start before COVID, but because of Right. Yeah. So I only started

slow travel day.

I already started last year Okay. In Hikari, Rwanda. Right. Right. So And

what is what is it you do when you get go over there? It's, like, a couple weeks, you said. Right?

We did the 2 weeks. Yeah. Yeah. So each, teacher taught 1 module. Mhmm. So some people taught the electronic structure. Mhmm. I did a small module about machine learning, yeah, presenting my research. Cool. And this year is gonna be Nigeria? Yeah. This year, yeah, we're going to be Nigeria, and next year this year, we call it a mini assessment. Okay. Yeah. Only maybe 30 people focusing on Okay.

Oh, cool. Cool. And this year, you're you're bringing your topic is gonna be quantum computing or something related to quantum computing?

Yeah. This year, I'm planning to teach student there some basic Excellent.

Yeah. That's great. That's really great. And what sort of you said mini is 30 people. What's the sort of normal size then?

Normal size is about double or triple. Right. Yeah. Double or triple. But much more countries still in charge.

I see. So you're bringing people to one location, but they're front coming from multiple countries. Got it.

So that means the 10 Africa or more than 10 African countries. Great.

That's amazing. Were involved. That's really great. Well, fantastic. Thank you so much for joining me. This has been a really good conversation.

Thank you so much, Sebastian. You too.

Okay. That's it for this episode of The New Quantum Era, a podcast by Sebastian Hassinger and Kevin Roney. Our cool theme music was composed and played by Omar Khosstah Hamido. Production work is done by our wonderful team over at Podfi. If you are at all like us and enjoy this rich, deep, and interesting topic, please subscribe to our podcast on whichever platform you may stream from.