Science Fair Series: Demystifying Materials Science

Kristin Persson and Gerbrand Ceder, co-founders of the Materials Project and pioneers of AI-driven materials design, are fundamentally changing how new materials are discovered. Join us to explore how computation and autonomous laboratories are revolutionizing materials science across energy, electronics, and sustainability, and to envision what it will take for the United States to lead a new era of scientific and technological discovery.
 
The Science Fair Series is a new meeting series highlighting cutting-edge developments in emerging technologies that will impact foreign affairs. Our series is specifically designed to bring scientific insights to all members of CFR, regardless of members‘ scientific background. This event is made possible by the support of the MacArthur Foundation, Rockefeller Philanthropy Advisors, and the Hewlett Foundation.

TAYLOR-KALE: (In progress)—on “Demystifying Materials Science.” The Science Fair Series is a new meeting series here at the Council, highlighting cutting-edge developments in emerging technologies that will impact foreign affairs.

My name is Laura Taylor-Kale. I am the senior fellow for geoeconomics and defense here at the Council. And I will be presiding over today’s discussion. I am absolutely thrilled to be able to introduce the Council to this distinguished group of scientists that we have here today.

So I will ask each of them to introduce themselves a little bit and explain a little bit more of what they do to give it proper justice, but to my immediate right here is Kristin Perrson. She is the Daniel Tellep distinguished professor of materials science and engineering at UC Berkeley, and faculty senior staff scientists at Lawrence Berkeley National Lab. She leads data-driven materials design efforts that integrate first principles, computation, high performance computing, and machine learning. And she’s the founder and director of the Materials Project, which provides open access to computed materials data and analysis tools to accelerate materials discovery.

And to my far-far right is Gerd Ceder. He is the professor of materials science and engineering at UC Berkeley and senior faculty scientists at Lawrence Berkeley National Laboratory. His research combines theory, computation, machine learning, and experiments to design materials for energy storage, especially next-generation batteries. And he is the cofounder of the Materials Project, along with Dr. Persson. So, again, delighted to have you all here today. Thank you so much for coming to Washington to be part of this Science Fair.

So we’re going to assume that this is a very nontechnical audience. I know that there are people in the room who have expertise in AI and also various other scientific areas, probably also in virtual space as well, but let’s really assume that this is a nontechnical audience. And what is the one idea that makes materials science click for you? And what is it that distinguishes materials science from, say, chemistry, physics, or engineering? We’ll start—maybe we’ll start at the at the far end with Gerd.

CEDER: You know, I used to call it physics for the future, but that didn’t go over well with my colleagues. (Laughter.) What I like about materials science is that it’s sort of—you know, it’s scientific. It’s physics and chemistry. But it’s with an engineering purpose. And the things you make when you develop new materials are—they are true enablers, right? You see other people, when you make a materials, they run with it and they build new technology. They build new things for society, whether it’s batteries or fiber optics or engines we can fly with. So I like the enabling part of it. It’s really fascinating to me.

PERSSON: Well, I don’t think any kid grows up thinking, I’m going to become a materials scientist. I’ve never heard that. I think it’s something that you learn later on. And like, Gerd said, it’s sort of a blend of physics and chemistry. But the one thing that sort of clicked for me early on, and I like history, and there’s a reason we name our ages after materials, right? The Stone Age, the Iron Age, the Bronze Age. And I think some people are claiming that today we’re living in the silicon age or the petroleum age, depending on which materials you think is more important.

And imagine all the things we’ve done with steel, right? Without steel, we couldn’t build the large buildings we have today. We wouldn’t have the Industrial Revolution. Without understanding fiber optics like silicon or defect-free materials, we wouldn’t have the information age. We wouldn’t have internet. We wouldn’t have any of the devices you look at. And without devices, without batteries, we wouldn’t have any of the cellphones ability to go off grid or, like, talk to people across large distances. So, yes, enabling, but also—and truly across the ages, from the very fundamentals to more advanced technologies.

TAYLOR-KALE: I also—it strikes me that there’s been quite a lot of advancements in materials science over the years. And we talked a little bit about this earlier. I’m just curious, of all the people that are here in the room at the Council, how many people are wearing, say, an Oura ring, or—(laughs)—right, exactly, some sort of a wearable? Ten, fifteen years ago the idea of having, you know, a wearable, you know, tracking your biodata was still a little bit of science fiction. It was something that, you know, we had in the movies, you know, Marvel movies and things like that. But now it’s very much reality. What are some of the other areas that perhaps we’re not aware of and making the link that materials science has really contributed to these new technologies and new things that we have, that can make our lives different and better?

PERSSON: So maybe you’ve heard—you’ve probably heard about quantum computing, right? And I think most people think that, well, quantum computing is going to be great, first for security, for being able to do problems we can’t do with normal computing. But I think it’s not quite as appreciated—and I talked to a lot of people who develop quantum computers. I don’t, myself, but I talk to them. And they are underestimating how much the impact the materials that are the actual qubits—(laughs)—that are part of the quantum computer.

There’s a lot of focus on the algorithms. There’s sort of the error-catching algorithms and all that. But in the end, if you don’t have materials that are going to do what you’re asking them to do, and long coherence times and the various sort of defects-free materials that doesn’t exist in reality—that physicists can come up with, but they don’t exist, you can’t make them—that’s one very current materials-dependent, critically materials-dependent technology. We will not be able to make it work on the scale we want to without coming up with better materials. And we don’t have them today.

TAYLOR-KALE: Hmm. And when you say “materials,” can you can you drill a little bit more on what we’re talking about when we say “materials”? So, you know, are they chemical components? Are they, you know, physical? What are we talking about when we’re talking about materials?

PERSSON: All right, so you have to stop me when I get too nerdy about this, but OK. (Laughter.) So imagine lots of, like, atoms, right? Lots of balls. And a lot of them inorganic. We call them inorganic materials that Gerd and I work on. They’re things like just simple aluminum oxide, for example. And that is one of the materials that you could put in a superconducting qubit. It literally contains a very thin layer of aluminum oxide. And when I say aluminum oxide, and I mean a material where you have aluminum atoms in certain places and oxygen atoms in certain places, and depending on where you put those atoms in, what arrangement, you’re going to get different physical properties of that material. And that’s what we call a single crystal. That’s when all the atoms sit in their perfect places.

But when we talk about materials and materials sciences, we also realize that that’s—very few materials exist as single crystals. They have, like, you know, this part here it looks like a(n) aluminum oxide, but there’s defects in it. And then there’s another crystallite next to it. And there’s some disordered stuff in between. So that’s what we call the higher length scale problem. And then there’s all kinds of composites and alloys and things in between.

CEDER: Yeah. I think, again, what makes it so fascinating is that the materials universe is more infinite than you think. As Kristin said, there’s a discovery aspect to it, where somebody comes up with a new compound. And in that sense, it’s parallel to chemistry. But materials are often the integration of multiple compounds. And that’s where the cleverness of scientists and engineers come in. You know, the reason your car tires are now—are so much better than thirty years ago is because people figured out how to blend the rubber with steel cord, which keeps the tire together even when you have a flat tire. This is why you can run flat tires, for example. So it’s not like somebody invented just a new compound, it’s because what—materials is also the combined engineering of things. You know, you could call them composite, but sometimes they’re much more complex than that. So what’s fascinating, it’s really an infinite universe of possibilities. And which is also why there’s so much work to do.

PERSSON: Concrete. I’m sure you’ve had—you know, looked at the rebar in your concrete, and understanding that the rebar is made of, you know, steels, right? But the interaction between the concrete and the rebar is crucial to the—how the thing is actually stable. And then the corrosion of the rebar, if you’re thinking of the longevity of that building, is a real issue. We have a daughter that’s a civil engineer. All she talks about is the corrosion of rebar. (Laughter.) Fascinating dinner conversation. (Laughs.)

TAYLOR-KALE: So there’s one aspect of materials and materials science where we’re talking about the ingredients, and sort of another part we’re talking about actual sort of bigger pieces, or systems. Can we talk a little bit more about the bigger pieces, the systems and the processes that go into materials science and materials research? We were talking a little bit earlier about—we’ll get into this a bit more when we talk about computation, of course—but, you know, the importance of the imperfections and the ways in which scientific research—you learn a lot from the things that you don’t—that you miss.

And one of the analogies you aptly gave was, you know, it’s like having a cooking show—(laughs)—and only showing the ingredients and then the final—the final dish, without understanding all the pieces, all the processes that go on in between. So can you talk a little bit more about some of the things that you’ve learned from, you know, that the middle part of the of the cooking show, the imperfections, the failures, the blow ups, and things like that?

CEDER: I think that, you know, one of the reasons materials has been so challenging for industry is because of the missing middle part that I think it’s traditionally been somewhat empirical science. I would say lots of, you know, skills of the trade. And because of that, it has led to extremely long development times for materials, right? So there’s a study that was done quite a number of years ago, when I was still at MIT, when the average time to commercialize materials is eighteen years. So that’s from concept to actual commercialization. That means some of them are out there at the twenty-five-year, twenty-eight-year scale. And it’s sometimes because while we understand the basic concept of compounds, once we start scaling it we don’t understand the role of the imperfections, we don’t understand all the things we need to control to actually make reliable products.

And that’s what we’re slowly starting to see changing. That’s in part driven by better science, but also by more rapid experimenting, by some use of AI. And honestly, that’s really the—I would say, the central challenge of materials. It’s not just the discovery, but it’s shortening that timeline for materials deployment. Because the problem is with a timeline that long nobody wants to invest in materials. That’s the one of the center—and nobody wants to do anything new anymore because, you could be lucky, and in ten years you’re done. It could be that you’re unlucky, and it takes you twenty-five years and you’re still nowhere, so.

PERSSON: We’re pretty close to Palo Alto and Sand Hill Road, if some of you know that. And if you go down and say, I have a new material, they look at you. No, no, no. We want software. (Laughter.) We can make money on software in two years. Materials take forever, and we don’t know if it’s going to work. (Laughter.)

TAYLOR-KALE: So let’s talk a little bit about computation, and artificial intelligence, and sort of data. We talked a little bit—you know, your whole lab, you’ve really advanced machine learning and the ways that you use data. And what do you think we can responsibly see through computation alone? And what is it there we actually need more experiments, and particularly materials sciences to step forward?

PERSSON: Definitely. So I’m sure you’re all using some LLM these days, right? And, you know, they’re pretty knowledgeable, right? You feel like that.

CEDER: You can admit it. (Laughter.)

PERSSON: I use them all the time. But the LLMs are trained on all the text that’s out there. And we, unfortunately, in the scientific world, what we publish is not all the data. It’s not all the things that we find out. So a lot of the stuff, when a researcher goes through an exploration, that person usually fails like fifteen times, and then comes up with the sixteenth time, and that’s the one they publish. They don’t publish—they don’t mention the other fifteen things goes into their brain, and they keep all those things. And that’s what we call experience, right? And that’s great. (Laughter.) But we want that, too. If you train any kind of machine learning or AI, you need the failures as much as the successes. So those LLMs that we have today, they’re only trained on scientific successes, by and large. So they’re very optimistic. Everything works. This is great. I asked, like, this new aluminum oxide with some defects or whatever in it, is this going to be a new superconductor? Absolutely. Good idea, Kristin. (Laughter.)

So we need more—

CEDER: Let me tell you more about this. (Laughter.)

PERSSON: Yeah, exactly. Do you want to hear more about it? Which defects should you do? Anyway, so what we decided to—we have to start valuing data not just for discoveries, but also just for the data itself. So about fifteen years ago we started what we call the Materials Project, which computes materials properties from Schrodinger equation, from quantum mechanics, which we can do today. We couldn’t do it when I was young. But, like, fifteen, twenty years ago suddenly we had good enough computers, and we had curated software that could solve the Schrodinger equation, that was finally ready to actually calculate materials properties. And we could do it much faster. And we could measure them using experiments.

So that gave rise to what is now today the largest and most used database for materials, properties and materials—sort of, various crystal structures, various compounds, all the different things. And if you pull in that kind of data, now suddenly these LLMs, these agents, these AIs can become a little bit more knowledgeable about not just the ingredients and the end omelet. Or if you look at it from a map perspective you’re not just aware of the valleys, you also see the mountains in between. So it’s the beginning.

But I will also argue that computations is not everything, because I can dream up any material I want in the computer, and I can calculate properties about it, but what if I can’t make it? It’s kind of useless, right? So we’ve had lots of people in the last five, ten years getting very, very excited about the AI ability to dream up new materials with fabulous properties. But then you go and talk to your friendly experimentalist, one sitting right there, and he looks at me, like, no. I can’t make that. (Laughter.) Tell me how to make it? And there is no Schrodinger equation for things like synthesis, because that is not just one equation.

CEDER: And I think if you think of AI in science, it’s maybe fundamentally different from your experience with LLM and AI applied to pictures and photos and texts that—in the universe of language, you know, a lot of word combinations have been made in the world, right? And that’s why LLMs are so good, and they can give you paragraphs that sound so real. They maybe can’t really write a good book, but because there’s just not enough good books. But science is very different. In science, most things actually have not been done yet. And so what you train on in science is a limited set of information. And the rest—either it would have to know all the fundamental equations, which we don’t know which they are. But the message I want to leave you with is that in science most things actually remain to be discovered.

We are at a—you know, if I had to put a number on it, if I could come back 300 years from now, we would probably go, like, well, yeah, we knew about 5 percent back there when we sat in that room. And 5 percent of the clever engineering that had been done on materials. So you live in a different universe where an AI needs to—it has no way of learning everything because it’s just not out there, right? Whereas with text, there are billions of texts of the English language, so it has such an ability to learn. That’s not true in science. And that means that the way for AI to be successful in science is to couple it to the physical world, right?

And so the thing I work on, in complement to what Kristin does, is essentially you build autonomous labs where robots do the work. They actually do physical experimenting. But AI drives them, so that now AI can do hypothesis generation and actually request the experimental validation. That also deals with things, like, you know, you don’t have to worry about hallucinations, because, you know, if it has a bad idea, it’s just going to request the experiment. And the data is going to come back and say, no, you were wrong. And so, again, the difference with AI and science is that we will not be successful without that coupling to the physical world, where AI drives, essentially, physical world experimenting, because we just don’t have enough information.

PERSSON: And it’s a little sad, to be honest, because we’ve been making experiments and doing experiments for a hundred years in materials science and in physics and in chemistry, but we only saved some—a very small fraction of it. So now we actually have to go back and do it with robots, because robots will save the data. Our graduate students refuse. We actually tried for many years. We tried to get the graduate students so please record your failures. So they looked at us, like, really? No. There was no way we could do it.

CEDER: But this is an important part, because what we collected, historically, of the scientific record was the data interpretations and the knowledge learning. And that’s probably exactly what the AI didn’t need. It needed the raw data because it could do its own learning from it. So in some sense, we’ve been poorly set up in science to use AI. And this is what we are now starting to see change. People spend a lot more time on, like, let’s record everything. Storage is cheap. Everything is electronically connected. Let’s record everything. Let’s do as much as we can, whether it leads to the final solution of the experiment or not. And that is the real fundamental change that you will see in science and AI driven science.

PERSSON: Have you ever had a grandmother who had a recipe for a stew that you couldn’t reproduce? (Laughter.) I did.

TAYLOR-KALE: Stews, cakes, all of it.

PERSSON: All of it, right? And you really felt at some point, I’m going to take a video of her as she does it, because she was unable to tell me exactly what kind of spices in what order and whatever she did with it. Whenever I tried, it was never was good. So it’s the same thing with materials, right? It’s like an experiment. Unless you were to take a real, like, catalog, pretty much everything, you would not be able to get all that metadata or provenance, the way we call it. But we’re getting better about it because robots, again, they will catalog everything.

TAYLOR-KALE: So this is where I wonder, is it a materials problem or is it a process problem? Do we not understand—so, you know, going back to the to the analogy with, you know, the grandmother’s stew. Is it that we don’t know the ratios for all the ingredients that she put in, the materials? Or is it something about the way in which she does it, her process—

PERSSON: The heating profile.

TAYLOR-KALE: The heating profile. You know, the temperature in the kitchen, the humidity. Maybe she likes to keep the humidity really high in her house, and so that affects her cooking. I mean, that is actually a thing in baking.

PERSSON: Yeah. For the rise. (Laughter.) Right? I like cooking.

TAYLOR-KALE: So, you know, as, you know, taking this step in thinking about, again, materials science versus other parts of science or even engineering, what is it that you are now discovering now, through the AI models and through machine learning? Is it actually a materials issue? And what is it that’s actually a process issue? Or is it you’re just trying to record everything, because we just don’t have all the information we need?

CEDER: I know it will sound like a cop-out, but it’s really both, right? So we do often—we do new compound discovery. And that’s often the start, kind of, a whole genealogy of kind of new materials that come out of that, right? So, you know, ten years ago I discovered a new compound for energy storage. But what comes out of that is, through variations in processing, a whole collection of slightly different materials that some will work in a battery, some will not—and, again, that’s coming back to maybe my introductory point, that materials science is this combination of physics, chemistry innovation, but then process engineering on top of that, right?

You know, if you—I don’t know if anybody has Nike shoes. This crowd is too fancy, right? But like, so—(laughter)—like, you know, if you go to Nike, right? All they do is—they work with the same rubbers. They work with the same, like, reinforcement. But, you know, they go, like, and constantly change the way they blend them, the way they shear them together, so that the sole has a little more elasticity or feels a little lighter. They don’t invent new materials. They invent new processing of existing materials. So there’s a bit of both going on. Industry is typically more on the let’s take something that exists already and modify how we handle it. We are often much more on the side—we are sometimes radically inventing novel compounds that then have to be processed.

PERSSON: And you can screw up a good compound with bad processes.

CEDER: Yes, very well.

PERSSON: Kind of hard to process a bad materials into a good one, but there might be examples of that too.

TAYLOR-KALE: So let’s switch a little bit. We’re in Washington, D.C., the center of the policy universe—from my perspective. (Laughter.) We have become very seized with supply chains and supply chain shortages here. We’ve been, you know, interested in batteries and how we can get next-generation batteries. We’ve talked about—lots of conversations in Washington happening around critical minerals and critical materials. And I’m really curious, especially, Gerd, since you worked on batteries and battery components for a long time now, what are some of the tradeoffs that you think dominate the next-generation of batteries that we’re looking at? Is it energy density? Is it safety? Is it cost? Is it supply chains for certain materials, like lithium and other things? What are some of the real tradeoffs that are dominating our ability to generate new batteries?

CEDER: The supply chain certainly is important in the conversation of new materials and batteries. And it’s—you know, not only because it’s a supply chain issue, but it’s also because supply chain issues often means higher costs, right? So as long as we work with things like nickel and cobalt and lithium-ion batteries, that a lot of it comes out of countries we don’t really like to work with or really can’t easily control. So it’s both a supply chain issue, but it also adds to the cost. Today raw materials cost is almost 50 percent of the cost of your lithium-ion battery. So that’s not a long-term sustainable issue. So we do try to engineer with materials that are, you know, not just available in the U.S., but broadly available, and lower cost. That’s, sort of, one of the issues.

There’s certainly a part of the energy storage community that focuses heavily on that. Sort of both for current lithium-ion technology, making inexpensive cathode materials which is where most of the supply chain issue is, it’s but then even trying to go away from lithium, right? So there are people working on sodium-ion batteries, so we don’t even have to use lithium. There are people working on very exotic technologies, potentially.

TAYLOR-KALE: Hydrogen. We can talk about that.

CEDER: Hydrogen. There you go. So—and then people work on other aspects that maybe they don’t quite worry yet about supply chain issues. Things like solid state batteries for higher safety, higher energy density. So supplies is certainly in the mix. But as you know, right, supply chain issues is a complicated issue. It’s not just about mining more, and there’s a whole lot of other aspects to that, that I think the battery industry worries about. But the battery industry is a big dog these days. I think people don’t realize but, you know, we’re not—more than a third of the world’s nickel production goes to lithium-ion batteries, which is insanely high. If you told somebody that ten years ago, they would have called you crazy.

TAYLOR-KALE: What worries you more about our energy and battery dependence? Is it the geological constraints or geopolitical?

CEDER: I think it’s more the geopolitical ones, right? So obviously we don’t have a manufacturing base for batteries in this country. And we—the issue is not so much that the supply of the raw minerals is—some of it is in countries we have friendly relations with. But the refining is often done in other countries, right? For example, you know, lithium is really kind of easily accessible to us. The largest producer is now Australia. But Australia doesn’t actually produce lithium precursor for lithium batteries. China does, right? Australia ships almost all their production to China.

And so it looks like China is the largest producer of lithium carbonate and lithium hydroxide and the other precursors, but it’s because Australia has chosen not to up-process, right? So in principle, we could do that. There is really nothing magical about it. We could be—the United States could be a lithium carbonate producer. You just have to send the ore ship from Australia to California, I guess. But maybe not in California, but you could cross into Nevada and Texas. But then you have to make it through the Panama Canal, right?

PERSSON: There’s also a CO2—and you have to deal with CO2 and the processing.

CEDER: And, you know, there are issues, right? Like, you know, you don’t want the CO2 balance on your record. That’s the other thing, right? So a lot of—a lot of minerals processing for the battery industry has become a lot more CO2 intensive than it used to be. And that’s because, especially for nickel, Indonesia has access to much lower-grade ore than they used to. So they’re going deeper and deeper and deeper, basically, right? So the ore is lower and lower grade. So up-processing that takes a lot more CO2. And that’s now on China’s books, not on ours.

TAYLOR-KALE: So, Kristin, as I think about the many conversations around critical materials, whether it’s semiconductors or, you know, rare earths, critical minerals, there’s also this sort of idea that there’s this promise that AI, and even quantum, have in potentially helping to leapfrog over some of these—some of these challenges. And what are you seeing from your standpoint and from the work that you’re doing in machine learning and AI with materials? And how realistic is it for us to expect that we can leapfrog over some of these challenges?

PERSSON: I certainly believe that’s the future, but it’s going to take some investment. As I alluded to earlier, you have to—no AI better than the data you trained it on. And if you train it on very select data, or incomplete data, or unsystematic data, you’re going to get very poor machine learning or AI performance. Imagine how many—you probably, hopefully, have all seen Iron Man, right? So if I say J.A.R.V.I.S., you know who I’m talking about? OK. So imagine J.A.R.V.I.S. being just as eloquent and beautiful British accent as he is in the Iron Man—(laughs)—sorry—but he doesn’t actually know much. That’s what we’re dealing with today. We’re actually pretty close to having an agent we can talk to that has a whole conversation with us. That we say, come up with a superconductor for me, and can you substitute this materials, this element, and that? Absolutely. J.A.R.V.I.S. will do that, and will be very eloquent talking to me. But he won’t actually know anything much, because that knowledge just isn’t in his training.

So we have to get the training data from simulations from the physical world, being held accountable by the physical world. This is a material I can actually make. Or maybe I can make different kinds of substitutions in this direction here, but I can’t make them here because then it falls apart and makes other materials. So they’re all those. And building those components together will be crucially important. We think of it as sort of active learning that our AIs are not—they won’t be—they won’t know everything on day one. They will know some amount, but they will go out and they will fetch that information. And that can only be done with automation, robotics, AI, different kinds of agents, trained on different kinds of data.

But as both Gerd and I know, building that infrastructure is not for the faint of heart. I’ve been doing it for fifteen years with the Materials Project, and that’s just a simulation aspect. So now we’re integrating the Materials Project with A-Lab, which is this automated robotics lab. And is an enormous amount of work, especially for academics. So and now imagine that all your industry needs to do this too. GM needs to do this. GE needs to do this. Corning is already doing it because they actually do invent new materials, which is darn cool. But all of—if we are industry competitive, we can’t just incrementally improve existing materials. We have to leapfrog. And it will do it, but we have to invest in that future. That’s my feeling.

TAYLOR-KALE: So what is it that government needs to do? What is it that policymakers need to think about in terms of creating that future and allowing that future to take shape?

CEDER: You know, I think that—I mean, obviously there are the easy answers, right? As scientists we should say there should be more money for science. (Laughter.) But I’m going to try to be a little—I think that, you know, given the facts about AI on the ground now, I think government needs to rethink how it runs science, in my opinion. It has—you know, we have an enormous science budget. Even though it has declined, it’s still an enormous science budget. We have enormous resources with great universities, the National Lab system. But I wonder if the old model of a scientist in the lab is over. I think that the—because, as Kristin was hinting at, that relied very much on apprenticeship, where the things that are in my brain that I learned the hard way get transferred to my graduate students.

And this is not how AI will transfer knowledge, right? AI will accumulate knowledge. And it just, you know, the next user will use that knowledge without having to do the apprenticeship. The other thing is about, as we get into robotics, right, you know, there are two ways to do automated labs. And essentially, speed is one of them. That’s the obvious one. But autonomous labs are directly addressable by AI. You don’t have to go through a human. They can communicate with other labs across the nation. And so, you know, is this model where I’m a professor and I have my three graduate students and I spend most of my time writing $100,000 grants, is this really still the way to go? Because we have historically grown into—this is the after World War II model, right, that we grew into. And it was very successful. But I’m not sure that we have thought yet about if we are ready to reconceive that model.

I can tell you that the academic world is not ready for it. I think that—you know, I don’t think the government’s ready for it, either. But I think the academic world is—I think this the first time in history that academic world is the incumbency. And, you know, that is a new role to be in, right? And we have argued with our colleagues. And it’s not going well, so. (Laughter.)

PERSSON: We need your help. (Laughter.)

TAYLOR-KALE: So disruption happening in the academic space with respect to how research is done.

PERSSON: And figuring out the role—because there is still a role for human invention, innovation, and creation. We see it, right? It’s not like my entire lab is run by robots and humans just walk around like, ah, I’m just a grunt, you know, coding all the agents. No, no .They are always humans in the loop. And they’re benefiting from the fact that the agents are running twenty-four/seven, and coming up with hypotheses, and then the human goes, like, very cool. You tried that, but I want to try that now. So they get—they get to leapfrog, basically, like, you know, based on the experiments, virtual or realistic, that the agents were doing when the graduate student was sleeping.

So there is really—I would say it’s going to need, kind of like this country invested into light sources and big equipment at some point. Then we keep investing into that. We’re now going to probably have to invest into these sort of automated, agentified resources, databases and robotic labs. And we’re going to have the scientists submitting ideas and interacting with them on another level than we used to.

CEDER: If I may add something to this?

TAYLOR-KALE: Yes.

CEDER: So I think what’s important, right, so we have seen AI being very good at certain things. It’s certainly better than humans in data accumulation, right? And AI can keep millions of things in memory. It can, within a millisecond, search through 500 research papers, right? So things that we cannot do as humans, right? Just to put it in perspective, every year, there are multiple millions of research papers published. In every given area, there are probably fifty to one hundred per day. So we’re already living in an age where we cannot absorb the information. So we’re already being inefficient. Where I think—what Kristin is saying—that when you make the tools better to, I would say, stimulate people’s creativity, every time we have seen that work. So if you get rid of the grunt work, what we have seen is that people actually get better. They, like, step up to the challenge, and they ask questions in a different way.

Because when Kristin and I first automated, we were among the first teams to automate computation in terms of what’s called now high-throughput computing. So basically rather than typing some input file and, you know, sending it to the computer, waiting for the answer we have—we would develop machinery and scripts that could do thousands and thousands of calculations without any problem. Just you set the command, do me these thousand calculations. And what we learned from it, it’s not like people just did the same work faster and went home at 1:00, right? They just started asking different questions of the science. And I think that’s what we will see, in my opinion, with AI as well. That people will actually become better scientists, because AI will help them and let them—give them parts of the answers that they will piece together.

TAYLOR-KALE: I think this is a perfect place to open it up to the audience for questions. So at this time I’d like to invite members here in Washington and online to join the conversation. Reminder that this meeting is on the record. And I’ll go to Mercedes first.

Q: Thank you. It’s Mercedes Bishop of Georgetown. Oh, thank you. Mercedes with Georgetown University.

A few questions. Please feel free to pick and choose which one is most interesting. I’m curious about the current centers of excellence for materials science, both here in the United States and also overseas. And, second, wanted to tease out a little more which tools would best benefit from U.S. government investments to stimulate creativity. Thank you.

CEDER: What was the first question?

PERSSON: The centers of excellence, can you elaborate what you mean?

Q: Yes. So, for example, I imagine Cal Berkeley—I imagine Cal Berkeley may be one of the centers of excellence for materials science, where you have the—

CEDER: Of course, yeah. (Laughter.) Goes without saying, right?

Q: My father got his Ph.D. in economics there also. Go Bears.

Yeah, if you could expand upon that. Other competition centers, because that oftentimes leads to the collaboration. I’m curious, what are the key collaboration centers of excellence? What would you like to strive to be for like—or, how it could this center of excellence at Berkeley be better, as well?

CEDER: You want to take that one?

PERSSON: OK. Well, in terms of the—

CEDER: You’re more diplomatic.

PERSSON: Oh, dear. (Laughter.) In terms of the—sort of the computational automation, if I take that first, because that happened first. So you may remember the Materials Genome Initiative that was launched by President Obama. That gave rise to my first funding, which was the Materials Project. And that sort of was the first computational materials database that has become as big as it is today. But, you know, our friendly neighbors, there was a center in Switzerland. There was a—there was a thing in Germany. There’s been a couple of them cropping up in China. So certainly, people realize that this is a good thing. And there’s been many other versions of it. I think still the Materials Project is still the most widely used across the globe. We are sending terabytes of data every week to the entire world, we are 4,000 new users every week. And I know if you’re Facebook or Twitter or whatever people are using now, that’s nothing. But in the science world, that’s amazing.

CEDER: Nobody uses Twitter anymore.

PERSSON: OK, fine. That’s right. Oh, I’m so dating myself. (Laughter.) So in terms of the computational aspect, I think, yes, there are certainly other centers of excellence. But in that place I will say, with some humility, that we are world-leading.

When it comes to automation of the physical, like automated labs, there’s Andy Cooper in Liverpool.

CEDER: The U.K. is very big here.

PERSSON: The U.K., very big on it. There’s some people in Korea doing on batteries.

CEDER: So what we see in Asia, especially in Korea and parts of China, is a lot of it is driven there by collaboration with industry. So there you see the autonomous labs being built, sometimes directly in the company, sometimes in the academic world, but almost directly completely sponsored by, you know, Postech, Samsung, you name them, all the big—all the Korean giants. So in particular, Korea has jumped on this dramatically, probably even more than China, I would say, at this point. I think I know of only a few autonomous labs in the United States, but Korea is, like, what, fifty, sixty million people, or something like that? And I know at least of ten, so.

PERSSON: EPFL also has—

CEDER: EPFL also—(inaudible).

PERSSON: (Inaudible.)

CEDER: So it is emerging. And I think in Korea what helps is there is—there’s still a much tighter relation between the academic world and the corporate world in Korea. For good or bad, but in this case for good. I mean, it works very effectively for them.

TAYLOR-KALE: Can you talk a little bit about how that differs from the United States? And then anyone—who has who has a question, just raise your hand so I can see you and know where the questions are. OK. Thank you.

CEDER: I think, you know, maybe somewhat frustratingly for us—so Kristin and I run a lot of—you know, we have a lot of government funding. We also have a lot of industry funding. I think at this point every single one of it is foreign. Every single one of them. So it’s Asia. It’s Europe.

PERSSON: I have one U.S. startup, but it’s Chinese funded, so I don’t know. (Laughter.) It’s in the United States, but it’s Chinese-backed. I don’t know how to call that.

CEDER: So I think our frustration with—I mean, and, you know, I’ll state that, you know, industry is not one monolithic giant, right? There are—companies have enormous differences in how they operate. But I think there are sort of two basic problems. One is that, you know, they are not that willing to invest a lot in early-stage research, even when it’s important to them. And, you know, that’s known. But I think the other one is that in—certainly in some companies, not all, research has been slashed so deeply that they don’t even have what I call receivers anymore.

Like, they don’t even have people who can receive science, in the sense that even if I did something great for them, they wouldn’t know because it’s been—they don’t have the people anymore who could understand the opportunity for them, right? Because a research product is never a finished product, right? It’s not like, oh, hey, I have a totally new steel for your car, and then you can buy 10,000 pounds of it tomorrow, right? It’s more like, oh, I have found a new way to engineer the microstructure to have a more ductile, or more—you know, better crumple zone steel. But you now have to work on scaling and figuring out how to—

PERSSON: By the way, I have one gram of it.

CEDER: Yeah, or whatever, right? So—and in some companies we don’t see the receivers anymore. And this is different. There are still companies that do this much better. But, yeah, it’s a frustrating situation, I would say.

TAYLOR-KALE: So I’m going to take a couple of questions, way in the back, and then here in the front, and then I’ll get to the other two afterwards. Thank you.

Q: Hi, Molly Ariotti. Former academic, but political science, and now I work for the government.

So I was thinking about this from the kind of funding environment of the current administration. And I feel like you’ve kind of touched on that already. So the back half that I’m interested in is whether you’re seeing shifts in academia in terms of how we publish, how we incentivize publishing, to your point about generating a complete universe of data, right, not just the selection on publication bias. Are you seeing those incentives start to change for tenure or for promotion in ways that you think will help address that? Thanks.

PERSSON: Oh, I really hope so. So I will say that we have, for—I mean, for decades we’ve been trying to get this to change. We do need to change the metrics in academia. We do. We cannot—

CEDER: But we will not. I mean, I think—(laughter)—

PERSSON: Well, there’s the hope. Anyway. (Laughter.) So, you know, publishing in Nature and Science is wonderful, right? But it is still a storytelling, rather than bringing science forward. And I still have some of my colleagues admitting that they don’t put all the secrets in the paper. They just want the—sort of the recognition, and then they’re keeping you holding back a little bit because, you know, it’s better that my group knows how to do this better than everybody else. So instead of being—

TAYLOR-KALE: Which means they’re not putting the failures in.

PERSSON: Yeah, no, absolutely not. The 15,000 different failures that happened before that. Which is really important to bring this forward. So no, we need to change to what is actually useful? What is useful? And today, a lot of the usefulness lies in data, and publishing data, and curating data, and featurizing data, and then, you know, metadata, and all those. And it’s really boring. Like, my students spend most of their time on this stuff. And I keep telling them, this is the really boring piece. But then you’re going to look at the data and do correlations and train in AI and do all the cool stuff with it. But you have to do this part first. And that that is a culture change.

TAYLOR-KALE: Anyone who’s suffered through a Ph.D. knows the pain of all the data work that you have to do. (Laughs.)

CEDER: So there’s—OK, there’s some positive sides. For example, you know, software is much more open, right? You know, when we have young students, everything goes on GitHub repositories. And so there is a certain openness on that. I think, on the data that hasn’t happened yet. And it’s—you know, so I was part of the Materials Genome Initiative under the Obama administration. You know, we went around. And we—I have—I have town hall trauma from this, where if somebody tells me, you need to hold a town hall, I run, and I run as far as I can. Because we have to sit in rooms with the scientific community and try to get them to agree on data standards. And it went nowhere. Nowhere. Absolutely nowhere.

We did not have the support of the government. They didn’t want to do it either. You know, this was fifteen years ago, but, you know, we think of this as a small error, but it was a giant error of the U.S. government. Because if we had fifteen or twenty years of collected scientific data—and not in papers, right—AI would be miles ahead today, because we would have, I mean, you know, the results of a few hundred billion dollars of research in data that AI can train on. And because of these mistakes we made collectively, we are nowhere. I mean, AI is so behind, right? Because we did not do this fifteen years ago. And even now, there is very little appetite for it, I would say.

PERSSON: Well, I mean, Genesis is a step in the right—

CEDER: Yeah, the Genesis is a good step.

PERSSON: It is, really. And you can see the excitement about it, despite the fact that Genesis is kind of underfunded for what it’s asking for. (Laughs.)

CEDER: There are 5,000 proposals being submitted, or something like that.

PERSSON: Five thousand proposals, or more, we’ve heard. So, yeah. No, and there’s a hope that this is the beginning of a change, and that we will stop thinking just in stories and thinking about usefulness in terms of what we’re bringing to the table, yeah, for the new generation.

TAYLOR-KALE: So, speaking of constraints, we have a lot of questions and starting to limit in time. So I really want to get to your question in, and then I know that there’s a question in the virtual space as well. So, if we—thank you.

Q: Tom Henneberg. And I have the pain of running a materials company. (Laughter.)

CEDER: Oh, I’m sorry.

Q: So you’ve said a lot of things that resonate. And I’m going to borrow your town hall trauma statement. (Laughter.) I think almost a decade ago a paper came out of MIT that estimated that in the prior four decades, which they refer to as kind of the information age, to your comment about ages earlier, that something like nearly three-quarters of all computing advances had been because of hardware materials, not because of software. But where that has gotten us in the ensuing decade, as we’re describing, is software is eating the world, right? Ninety percent of venture capital in the U.S. last year went into computing AI, right? You got to put AI on everything.

You just described something that I hadn’t arrived at yet, wonderful insight, that it’s going to make people more—researchers more creative, because going to free them up from other things. But it hit me when you said that, that we have also created maybe a more capital-intensive startup cost to get there. Because my company, we started relatively leanly. And we’ve got, you know, a tube furnace and some beakers and some mixers and a lyophilizer. And we got this stuff, and people run back and forth between them. If I had to set them up so that a robot could use all of their functionality and do—I’d be buying millions upon—I’d be buying tens of millions of dollars of robotics, that I couldn’t raise funding for because Sand Hill Road wants to fund software and AI.

So could you comment on this? Can you reconcile that we’re going to make people more creative, but we’re going to narrow down to only a few places that can be doing the research and afford to do it under their own culture, which, of course, limits creativity, which is the opposite of what you said? I’m going to be debating that in my head for twenty-four hours. Help me here.

CEDER: Well—

TAYLOR-KALE: Before you—before you answer, let’s go to get the virtual question, and then we’ll have the two questions circulating.

OPERATOR: We’ll take the next question from Cliff Rhoten.

Q: All right. Good afternoon. My name is Cliff Rhoten. I’m a portfolio manager at 3M Company.

And you mentioned kind of this inability of corporates to receive science. And I’m wondering if the opposite is true, and if the relationship is appropriate to pass on those lessons about scale up and process to the public institutions. Thank you.

PERSSON: All right.

CEDER: Dow do you—

PERSSON: Which one do we start with?

TAYLOR-KALE: Whichever one you want. Two very good questions.

PERSSON: How about we start with the last one?

CEDER: I want to start with the last one and come back to—

PERSSON: Yeah, yeah.

CEDER: Yeah. I think it’s a completely appropriate comment that it is—you know, it’s a two-way road jam, the relation between academia and companies in this country. It is very hard to get our colleagues interested in working with companies. I think that was historically a lot better. And in some fields, has been a lot better. For example, historically in metallurgy this was much better. You know, metallurgists worked with metals-producing companies. And I can’t really put my finger on how it’s become so bad. You know, Kristin and I have had very good relations with some companies that are willing to share things, and very poor relations with others.

PERSSON: I’ve learned the best things from working with companies. Really. And I’ve learned to be relevant, to work on things that actually will matter and actually today may make a difference. And I know—again, I’m not—I don’t want to throw any dirt on my colleagues. Some of them work on problems that may become important twenty years from now. And we need those too. But we also need people to work on things that are currently important. And that’s what I learned when I talk to companies. So I absolutely agree with what was asked. It’s a two-way street. Both directions need to change.

CEDER: And, oh, yeah. So the cost, yeah. You know, today, robotic labs are expensive. I mean, when I did it, you know, I was kind of thinking, this is the last thing before retirement. All the chips go in. (Laughter.) And it was—you know, I hate to use the word trickle down. I hope there’s a trickle-down effect. But it will get cheaper. I mean, we really are sort of building it from scratch, almost. We’re buying robots, programming them, coding them. But we are starting to have interaction with vendors.

For example, tool vendors, right? And, you know, we talk to a lot of these companies. We now do deals with them, where, you know, they’ll give us equipment cheaper, we’ll help them build an API for it, build automation for it. They won’t exactly copy it, but they usually get inspired by it. So, you know, I think the ecosystem will get cheaper, but it is not today. I mean, there are a few startups in the U.S. in this space. And they are all seriously well-funded. Like, seriously well. I mean, some of the numbers are insane. It’s sort of Sand Hill Road going completely out of control again. But yeah, I’m part of one. And if I see the money that shows up, I go, I could never dream of this. So there’s certainly a lot of excitement in that space.

PERSSON: But the hope is, of course, that they’re going to show some of the promise as well. And when that promise is shown, that hopefully that will also incentivize other people to invest in this. And the more we invest in it, hopefully the cheaper it will get. We have made a lot of mistakes around—just figuring things out, and how these different stations talk to each other, and how to program the robots. There will be OS systems to be bought that will hopefully be cheaper than having a graduate student work on it for three years and making a lot—so is it going to be only three places? I hope not. There might be—(inaudible)—places and there might be smaller versions, but then you can send off; just like if I want a really good electron microscopy pictures, I know where to go. But there’s also cheaper machines that you can get a fuzzy picture to start with. So my guess there’ll be a tiered system.

TAYLOR-KALE: So—OK. So I see three questions, one in the front, one in the back, and then here as well. Let’s go ahead and start in the front.

Q: Hi. Matt Merighi, X-Bow Systems. Maureen says hi.

I had a quick question about the—sort of where the Venn diagram overlaps between different elements of materials science. You know, to your quip about, oh well, you know, organic versus organic materials, and how different they can be. I work in solid rocket motors, where the materials science for the propellants, the liners, and the cases can be, I’m sure, entirely different. And then that’s nothing compared to the silk that would then coat, you know, pills that go into human bodies and whatnot. So particularly as you’re thinking about the next generation of materials scientists, you know, is the model going to be that you need to train materials scientists that can then branch out into these different fields of biotech, solid rocket motors, et cetera? Or is it more promulgating a knowledge and a mindset into different fields, like biotech or other physical sciences, to have a materials mindset to then be able to make sure that new materials are getting discovered? Because I’ve just—you know, when you go past the crystal level how different are those different disciplines? And how much will, you know, materials science truly be advanced in the applied side versus on the pure science and the pure materials side?

PERSSON: It’s a good question.

CEDER: OK. So I think I’m going to slightly disagree with you, if that’s OK, because I think that there is an inherent division between fundamental and applied science, which is, I think, slightly fake and has been made in this country. In the sense that, you know, there are fundamental problems in applied work, And I think maybe we have failed to see those and to educate people on those. But every applied problem I have worked on, you know, we—very often there are many pieces that you bring down to fundamental science aspects. And I’m going to give you an example of where it was done historically very well. You know, in metallurgy, this was historically done very well. If you look at the science of, you know, the 1930s, sort of, through the 1950s of metallurgy, and how people started understand processing of steels, making them better, you could think of that as that was very applied. I mean, it’s really what you do with the hot rolling mill, right, and the stamping of the steel. But it was treated at the very fundamental level. People thought of it as fundamental, but it helped the applied science go forward.

I think the mistake we often make is that we say, to make—you know, to make impact in applications, you have to do exactly what they do in the applications. You need the 3,000-pound rolling mill or you can learn anything about rolling steel. But I do think it’s a bit of a mindset that we need to relearn. And because I agree with you, at the surface these problems all look different. But I think when you get a really good scientist in there, they will figure out things that they transfer from other fields that will help your problem.

But, again, I think we may have to—you know, this comes back the same problem, right? It’s hard for us not to teach these things, because we’re often not informed anymore about these problems in industry, because we’ve lost that relation. So I can’t teach about the silk. I can’t teach about, you know, some of the rocket motor problems, because that relation has been lost a little bit. Whereas, you know, my original training was in metallurgy. I can teach people about—you know, even though I do atoms and stuff—I can teach about what hot rolling and stamping of steel does, and, you know, what fatigue means. And so I think we need to build that up again, because—yeah.

PERSSON: So civil engineering does that really well at Berkeley. They bring in people from industry and have them lecture. I think we can do that in more domains. And I was also thinking that, from the perspective of—we need—the one thing we are still going to need. You might think we all—oh, we’re just going to need software engineers in the future because they’re all going to code this AI, and the AI is going to figure out all the materials science. Personally, I think that’s completely wrong.

CEDER: The AI is going to write their code for them.

PERSSON: Right, exactly. (Laughter.) Yes. We actually still really need people that understand their domain. Because, again, there’s a lot of hidden data out there, both in industry and academia. And people are really good. So that, we’re going to need. And retraining somebody from a metallurgist to a polymer scientist, I think is less of a lift than retraining a software person to a materials scientist.

CEDER: Yeah. We still hire domain scientists and we teach them AI, not the other way around. Yeah.

TAYLOR-KALE: So in our last couple of minutes we have two final questions here. And I will also throw one in at the end as well.

Q: Hi, Alex Palmer, with the New York Times Magazine.

You had mentioned that one of the things holding us back is this older model of apprenticeship in graduate school that isn’t sort of built for what the AI age can make possible. Obviously, that’s a long-held system. There’s a lot of entrenched interest involved in it. What do you think it’s going to take to start breaking that and building a new model for American universities?

TAYLOR-KALE: Great. And then this question here.

Q: Hi. Meredith Broadbent from CSIS, a think tank.

Everything in D.C. is versus China. And what can we learn from how they’re doing things and organizing their research?

TAYLOR-KALE: All right. So since we are on our last two minutes—

PERSSON: Rapid fire.

TAYLOR-KALE: Rapid fire. (Laughter.)

CEDER: The entrenchment, I fundamentally—you can’t really change people. So the way to innovate is to build new institutions and get rid of the old ones. (Laughter.)

Q: Oh, creative destruction, good or bad? (Laughter.)

CEDER: Yeah, I mean.

PERSSON: Or, you know, carrots always work, right? If you just say, this is what we’re going to fund. We’re not going to fund this. Then the people will—the scientists are very good at—we’re metric driven. (Laughs.) So just tell us where you want us to go, and we will go there. We might complain, but we will go there.

CEDER: Are you taking the China question?

TAYLOR-KALE: And they, yeah, what can we learn from China?

PERSSON: Oh, the China question. Oh, well, longer funding models. Yeah, yeah. Because we are too much—like, we get jerked from one thing to the other because we have two-year funding and three-year funding. And we just need, especially for these larger projects, you know, we’re going to need longer—

CEDER: CATL has more than 10,000 people doing battery research.

TAYLOR-KALE: What’s CATL?

CEDER: It’s the big Chinese battery company, the biggest dog in the world now. So, you know, I have people there. They send me occasionally screenshots of my slides being shown on the screen. (Laughter.) You know, it’s like they’re doing everything, so.

TAYLOR-KALE: Well, final, final, final last couple seconds. Each of you, name one no regrets investment, whether it’s scientific, industrial, policy, that would accelerate materials-enabled, progress in your view.

PERSSON: Investment?

TAYLOR-KALE: No regrets investment.

PERSSON: From our—that we did? That somebody did?

TAYLOR-KALE: That the government could do.

PERSSON: Oh. Oh, well, I advocated ten, fifteen years ago for a national data bank of materials. And the Materials Project, if you think, that isn’t that. That is a core program. It has a funding cycle, and it could get killed tomorrow, if that’s what it is. The Protein Data Bank has been there since 1971. There’s a reason we have AlphaFold. Do the same for materials.

TAYLOR-KALE: Gerd, ten seconds.

CEDER: I was going to pick the same one. (Laughter.)

TAYLOR-KALE: OK. We have agreement. (Laughter.) Well, wonderful. This has been absolutely wonderful. At the Council we start on time and end on time. I really want to thank you both for joining us for this meeting. Thank the speakers, thank the participants as well, virtual and here in Washington. (Applause.)

(END)

This is an uncorrected transcript.