Transcript
Jonathan Godwin: Foundation model for material science means a model that stretches across the periodic table and stretches across the gamut of different types of materials. Semiconductors, batteries, catalyst materials, even protein materials, biomaterials. It can stretch across all of that.
Daniel Darling: Welcome to the five year Frontier podcast, a preview of the future through the eyes of the innovators shaping our world. Through short insight pack discussions, I seek to bring you a glimpse of what a key industry could look like five years out. I'm your host Daniel Darling a Venture Capitalist at Focal where I spend my days with founders at the very start of their journey to transform an industry. The best have a distinct vision of what's to come. A guiding north star they're building towards and that's what I'm here to share with you. Today's episode is about the future of material sciences. In it we cover the discovery of new materials using AI, carbon neutral compute, data centers built in space and the convergence of biology and materials.
Guiding us will be Jonathan Goodwin, CEO of Orbital Materials, a materials science company that leverages AI to discover and design new advanced materials, particularly for clean energy, carbon capture and data center applications. Founded in 2022 out of London, Orbital have raised over 20 million USD in funding from the likes of Radical ventures, Nvidia and Toyota. They've already developed the world's fastest and most accurate AI model for simulating advanced materials, including their own proprietary foundation model. The startup has also entered a multi year collaboration with Amazon Web Services to develop new data center decarbonisation and efficiency technologies. Jonathan's impressive background includes a tenure as a senior researcher at Google DeepMind where he honed his expertise in artificial intelligence. He holds advanced degrees in physics and computational sciences with a degree from University College London.
Daniel Darling: Jonathan, how great to see you today.
Jonathan Godwin: Great to see you too. Thanks for having me on.
Daniel Darling: Absolutely. So before we talk about how Orbital Materials is transforming material science, can you first ground us in how material science development discovery has being traditionally done to date? And what was the status quo before Orbital entered the scene?
Jonathan Godwin: Well, the development process hasn't changed for 200 years or so. The way that we were doing it back in the 19th century, the 20th century, it stayed exactly the same. Had a bunch of scientists who a huge amount about chemistry, a huge amount about material science. They have a lot of intuitive knowledge there. We use rules and heuristics and chemistry to reason about what the performance of a material is going to be. And then we go into the lab and we go and run a bunch of experiments and we see which materials have the performance characteristics that we care about and we're often surprised. We often see things we weren't expecting and those things that we weren't expecting and the ones that ultimately end up being really commercially valuable materials. So it's this really slow, painful process of trial and error and unlike so many other forms of scientific discovery or engineering, hasn't really up until AI being accelerated through computers.
And the reason is, is that the stuff that happens at that quantum level, the stuff with the weird metals like uranium or e.g., palladium or germanium, those sorts of quantum interactions that are going on there that give advanced materials these extraordinary properties, those are too hard for computers to model, they're too hard for us to think about as humans to do those calculations ourselves, and they're too hard for computers to model before AI. AI has really massively, it's been a zero to one of binary ability for us to model and understand the rules, simulate what's going on at that atomic level. And once you can simulate that, once you can understand that computationally, then you can start designing, then you can start doing rational design, then you can start leveraging things like generative models, LLMs, other forms of AI to optimize those materials. And that's what AI has really brought, which means that you can massively, massively accelerate the development process of new advanced materials. That's the kind of status quo and that's what AI is able to bring.
I think Orbital was really the first company to take this approach. We were working on very similar sorts of ideas back in DeepMind in 2020, 2021, 2022. Same sort of time as people were starting to think about large language models and we were seeing what was going on there, seeing what's going on in AlphaFold, for proteins and thought, well, materials is next. But to really bring this into the real world and to show that it really works, you need a lab, you need an integrated organization that has the capability to take these design tools and go and make them, fabricate them and bring them out there into the world. That's the kind of founding principle between Orbital Materials and why we started the company and how we're changing. I think that entire development process.
Daniel Darling: And I'd love to unpack that full stack approach you had, but maybe we just focus a little bit on AI because you really develop the world's leading AI model for advanced materials. And the claims are the outperform proprietary models from Google and Microsoft and can you explain about how you went about building this AI model.
Jonathan Godwin: And the reason why you want a foundation model in material science is because unlike I think areas of biology where you're only talking about things like organic molecules, only really a small subset of all the different types of molecules or materials out there in the world. When you think about material science and the scope of what we want to try and achieve, orbital we're talking about everything in the periodic table. Everything you could possibly throw at, your model. We should be taking like the weirdest elements in the world. How do they interact with water, how do they interact with CO2? They, what are their semiconducting properties? And so that means we really need that foundation, right? Visually you think of a foundation as something that underpins a huge amount of different applications, different work. And that's what large language models have been able to do. We can use them for coding, we can use them for copywriting and use them for for scientific research.
Foundation model for material science means a model that stretches across the periodic table and stretches across the gamut of different types of materials. Semiconductors, batteries, catalyst materials, even protein materials, biomaterials, that can stretch across all of that. And so that's what we did with Orbital. That was our primary goal, to get a model that can really encompass the gamut of where valuable next gen materials can be used. And that's what we've seen with the foundation models we built.
Daniel Darling: You break it down in kind of different buckets. You've got material discovery, you've got material design, and then you've got material development. So I'd love to tackle each one from there. So with material discovery, finding out what is the optimal material to use for a given project. When prompted with a desired set of properties, your generative AI, Linus reasons through the universe of possibilities of chemistries of the periodic table, to find what is the right material to use. So how does that actually work? And can you give some examples?
Jonathan Godwin: When you think about Linus and you think about a generative AI, you start with a functional property. So you can take something kind of simple, something like a new CO2, capture material. There are so many different ways in which you can construct these kind of sponge like materials. These things sort of come together a little bit like nets. You've got the metal as an anchor and then it's connected to another anchor through this long string like molecule. These create these kind of large nets which these nets or sponges go and capture that CO2 so you can Think can take any number of, there must be billions of organic molecules out there and there are hundreds of different types of metal. That combination is massive, it's huge. You can't possibly go and synthesize every single one of those materials. And it might even be computationally intractable to go and simulate each one of those materials and to get an answer even from a computer simulation. That's billions of computer simulations.
So you need a way in AI to search through this space, use chemical reasoning to hop here and to say, well, what the property is going to be like there, reason about those results and hop somewhere else and to do that scientific process for you to think up new ideas, test those ideas and then to generate a new hypothesis that allows you to get towards the correct answer. And that's what we do really with the generative AI. We prompt it with a functional property, a set of operational characteristics that we want the material to have. And then our generative AI will come up with suggestion. We can quickly assess that suggestion through AI accelerated simulations. AI can use that feedback to come up with another suggestion and iteratively work its way using its chemical knowledge to a really strong candidate. So that's what we mean by that kind of AI design phase.
Daniel Darling: How much of that is just independent of human involvement?
Jonathan Godwin: So we never go into the lab and make anything without a human there. So we really think about these tools, as accelerants for experimental chemists and material scientists. But actually I think being an incredible research material scientist or research chemist is more like being the world's best or one of the world's best violin makers. It's an artisanal craft that requires an unbelievable amount of time and manual dexterity and expertise in order to get to the results that you want. And so there we really think about these things as aids to that, that extraordinary scientist. So we can do a lot of in silico optimization hypothesis testing. But when it comes to actual lab testing we rely upon the decision making expression of our scientists in Princeton.
Daniel Darling: The next step is, once you've maybe discovered these novel materials is the design which is really leveraging a proprietary data that includes process, wet lab data manufacturing and performance factors. What happens as part of the design process.
Jonathan Godwin: When we think about the next stage of optimization, we'll have to take into account things like manufacturability and yield and things like that. And there we have a lot of own data that we can then use to do that optimization. But other forms of AI predictive AI tools that allow us to assess very quickly whether a new candidate is going to pass value sets of really practical and important concerns. There's I think one thing that's really different when you think about material science versus a drug company. But you also have a cost target that so cost target is a really important part of making a, commercially relevant and breakthrough material. It's not a breakthrough material unless you're hitting that cost to manufacturability. That was something that was in all of the great material science discoveries from the very best commercial labs, something that imbues those labs to the very bones, is that it's not an innovation until you figured out how to make it at scale, until you figured out how to make it at cost.
Daniel Darling: That's really that vertical approach now that starts to move into the deployment and production phase. And for that you've got a prototyping site in Princeton, New Jersey where you fabricate and pilot these products. Can you walk us through that facility and how you bring these discoveries into reality?
Jonathan Godwin: That facility was started at the beginning of 2024. It's a 5,000 square foot facility. We've got space for small scale commercial manufacturing there. We have our team, our team is research chemists, research material scientists, chemical engineers, process engineers, people who have expertise in not just the initial synthesis and testing but the scale up, estimating the cost models, building manufacturing, process chemistry. We're building out the competency there to really take everything from that initial discovery all the way to a full scale prototype that can be used for piloting and commercial demonstration. We've recently struck our first strategic partnership with AWS there. We've developed a material that we Orbital Materials is going to pilot. And we started the design of that material only in the early stages of last year. We had designed it, synthesized it, tested it and brought it to the crux of a strategic partnership in the space of a year. And Normally that takes 10. So that's part of the AI, but it's also partly just the incredible team that we brought to that site.
Daniel Darling: Let's dive into that because I think know you've outlined the process incredibly well. I love to dive into some of the use cases and what a great one to focus on with Amazon Web Services. And so what are you doing for them? Is it involved in their data centers?
Jonathan Godwin: So our strategic partnership with AWS covers materials and process technologies for data center decarbonization and efficiency. The thing that we're most developed in under that sort of platform that encompasses multiple technologies is our carbon removal technology and we're going to build that CO2 capture system. So we've got a proprietary hardware design that fully utilizes the unique properties of our material. And we've tailored that entire hardware design to utilize the waste products of a data center. Things like the waste heat that's generated from all of those GPUs and CPUs going. And we use that as basically the energy input, to do that CO2 capture. Because one of the things that we care most about in CO2 capture is why you're using all this energy. Where is that energy going to come from? Is that the best use of the energy? Well, it's a waste product, yeah, it's a great use of that energy. And we'll be piloting that at a data center. Get to be confirmed exactly whose data center, where that is. But we're working with AWS to closely monitor the performance of that system with the hopes that we can use that to decarbonize data centers and decarbonize AGI. So that's really exciting.
AWS is a big believer in decarbonization and carbon removal. So I think there's a shared recognition on I guess more traditional. How efficiently can I run this data center? How many GPUs can I fit in here? How can I reduce my costs and my energy usage? On that front end, materials have a massive part to play here and they were. The thing that we're most excited about at Orbital is new cooling technologies. As GPUs continue to become more and more energy dense, they become harder and harder to cool. And so you need new cooling technologies in order to facilitate the next generation of AI. Let alone the fact that you need to then reduce your costs as well because you need higher load cooling technology so you need more efficient ways of cooling. So there we're working on new thermal management materials and products to help scale out AGI and next generation hardware accelerators.
Daniel Darling: It's such a massive problem and so incredibly well timed for this new wave of technology from there. So you've obviously spent a bunch of time in some advanced data centers. How much efficiency can be brought to these data centers with new materials.
Jonathan Godwin: There are about 10 other things within the semiconductor manufacturing process that you need to really improve upon from a materials perspective in order for us to continue this AGI rollout for continued AI to get better to drive down that cost and increase the amount of compute.
So at the very core there in data centers you got the chips and the chips is clearly a material science problem. But if you kind of move you away from that because I think that's not what we're working on right now anything about more broadly. So the next big thing you think about from a cost perspective, next thing is actually really going to be the cooling, right? The cooling is going to be maybe up to 30 to 50% of the operating cost. It's almost about the same size as ah, the cost of powering the chips. That's only going to become more and more expensive. Instead of having these big fans, basically the pump air over the chips, you're gonna have to move to a situation where that cooling happens through pumping some form of liquid. So you move to liquid cooling and then you need different types of hardware and it's more mechanically challenging as well. So that expense comes up, but also just the cost because you're removing so much more heat because the chips, hotter, that costs more money as well. So I think you see that cooling and thermal management as a challenge becoming really, really huge.
So overall that energy cost becomes really, really massive. And if you want to do that in a clean and cheap way, because it's got to be cheap because the energy demands are so high, you need new forms of energy. And there you get to the third area of material science. One of the things that we have started to think about this is beyond five years for us, but one of the things that we think our current generation of models of being able to do that others would not be able to do is model things like what is the behavior of some of these nuclear fuels? And that sort of simulation is incredibly valuable when you think about designing new reactors because of course you have to do so much simulation computationally in nuclear because running experiments is so dangerous and so costly.
Daniel Darling: And so wrapping that all together, if you do fast forward five or so years out, what does a data center of the future start to look like to you?
Jonathan Godwin: So I think you'll start with that small nuclear reactor I think at each data center. Yeah. So this is, this is land, this is a land data center. And we've got some questions about space data centers. But I think it so like a data center campus while its own nuclear reactor. But when you go into a data center it will look very different at the moment you see these racks side by side by side by side and you know all these leads coming out of these racks and that's not what you'll visually see when you go into a data center. That's cool in a new way.
What you'll visually see is these massive vats, these baths. They've dunked all of the different chips into it these baths will be bubbling. So if you can make these baths made out of Perspex or glass or something, you'd see these chips and you see this bubbling water floating all around these chips. And that bubbling is what's causing the cooling going to happen. The phase change from that cooling liquid into a gases is using a huge amount of heat it drawing that heat away from the chips and keeping them cool. So that's I think what you'll see.
So instead of these kind of rack side by side by side you'll be going through what looks like these huge swimming pools full of a viscous kind of liquid that is bubbling away like it's been cooking. So that will look incredibly different. And you, you may not be able to see this visually. The complexity of the wiring is, can just continue to increase and increase and increase and increase and increase. So you'll see way, way way more wires going on. That's because you we'll be doing the scale of the requirements for fully interconnected training runs is going to increase, increase and increase. I
don't think that we're going to get to a point where we're decreasing the scale or the number of chips that need to begin interconnect it. If anything that's going to increase way more if we decide to move from larger and larger training trips to a larger number of smaller training trips for AI. So you'll see a massive increase in the complexity and cabling going on there. Those are two things that you'll see in the data center in five years time. So I think it will look very different.
Daniel Darling: Incredible visual, incredibly futuristic and it sparks in my mind something that's far more biological in look and feel. And obviously we've got this other part of your industry which is all around this bioeconomy or using sort of organic materials in material sciences. Is there a natural convergence in how biology starts to operate and how material science starts to develop our physical products? And maybe the data center is a really good example that you're starting to see.
Jonathan Godwin: I think one thing that's been really under invested in because of the challenges. I think some of the computational challenges have just been really huge and I think some of the syntheses have also been pretty new in developing this is kind of hybrid materials. So one thing that we've done as part of our outreach and scientific discovery efforts is try and shed some light on some of the mechanisms that go in the body between proteins and metals. Because AlphaFold’s done Nobel prize winning things change drug discovery. But one thing that it can't do is say what happens when you eat a banana and you get the potassium from that banana and what does it do in the body? Because it doesn't know anything about any metals.
We need potassium. Most of the cells in our body, most of the cells in all living organisms have a potassium ion channel which is something that basically says I'm going to let potassium through but I'm not going to anything else through. I’ll let potassium into the cells because I'm going to use it, but I'm going to let any nasties out. So it's like extremely selective. It's got this weird mechanism that's going on that selectively lets potassium metals into the cell. But no one can really understand it from a simulation point of view before because it's a metal.
You can't really run large scale simulations with things with metals in and none of the AI tools have previously been able to do it. So we ran some work, we did some work to try and understand the mechanism that's going on in that potassium ion channel. I think we discovered something that's pretty novel, we hadn't seen before in any form of simulation. We put it out. There was a research paper and it, I think a lot of brilliant people have been really interested and, and excited by it and I think that combination of organics and biology and some of the things that you get from traditional material science, we're at the cusp of those sort of things being computationally tractable.
And people have been working especially I think over the last 30, 25, 30 years in the lab on a lot of these sorts of materials. And I think that sort of area has always been an area that's been incredibly innovative. Hasn't reached necessarily commercial prime time yet, but with the aid of things like AI, I think will lead to you know, an explosion of new materials that kind of sit at that intersection of inorganic chemistry and material science and biology.
Daniel Darling: Incredible. And are you starting to see that from your researcher perspective where all of these people have been diligently researching prior to AI for decades sometimes on a problem and now they're equipped with either a new foundational model that has a really deep understanding of let's say material science like what you have, or something like AlphaFold that understands proteins, etc. Are you starting to see them break through those more intractable areas and start to really close the loop on that research and get to the discovery they'hope you to get to?
Jonathan Godwin: Yeah, I think that's happened in a bunch of areas already. I don't know, this story must have been told before. It was something that at the time internally at DeepMind when AlphaFold was announced, I think the team that was running the CASP protein folding competition was so amazed by AlphaFold 2's results that they wanted to double check that there hadn't been some form of cheating going on. And so they took a protein amino acid string that hadn't had a crystal structure associated with it, that people been working for about 10 years to try and find the crystal structure but just hadn't been able to resolve the experimental data. And they took this amino acid string and they asked AlphaFold to predict that AlphaFold predicted it. The team didn't know there was no structure behind it, gave it back to the organizers and then the experimental team that had been trying to resolve the structure for 10 years looked at the AlphaFold’s predicted structure and it was correct. They managed to then fit it to that data and it resolved that question about how to resolve that experimental data. We came up with a new protein structure that was an example to me of just how extraordinary this technology can be that can solve something that people have been spending 10 years trying to do.
Daniel Darling: And how long would that take? A couple of days for AlphaFold to figure that out?
Jonathan Godwin: That took I think a couple of minutes. And so I think we are in a different age of scientific discovery
Daniel Darling: We’re gonna come up on time here and theres so many different types of interesting topics we could go into. But I want to circle back to the comment you made about data centers in space and that maybe there’s some frontiers there and you do do things in aerospace and aerospace materials. But what is the fascination around putting a data center in space? And what is some of the cutting edge work you’re doing in Orbital?
Jonathan Godwin: Yeah, I think AGI has the capacity to be the most powerful thing that we ever discover. If we can get it right it is the most important thing to scale because more intelligence means more good. So we really got to figure out the physical problems that can prevent AGI from scaling and data center Rollouts is one of those. And we're going to run out of space on the Earth and we're going to run out and we're going to have less access to energy on the Earth. And so I think that what that means is that we've really got to figure out how to get a lot of that compute up into space. I think it's a really exciting and extraordinary thing to be working on from a material science perspective because once you're out in space, everything kind of changes.
So I gave you that example of these baths, these boiling baths where you put your chips in. That's almost exactly the way that you would also try and cool something in space, right? Because you can't use air cooling, there's no air. So you’re gonna have to bathe your chip in something and it's likely to be something that boils because that phase changes is really energy efficient. So you, know we need more cooling technologies.
Solar panels —the heat difference between the side that's facing the sun and the side that's not facing the sun is huge. And if you have really heat sensitive computing, material computing going on in what you're enclosing, then maybe you need to really think about what's that material science thermal management issue. And then you think about on the purely sort of communicating information and high bandwidth back down to Earth. There are going to be material science problems in those electronics as well.
Daniel Darling: Well Jonathan, you've done incredible work in just a couple of years with Orbital so no doubt in 5 years the transformation that this will bring will be incredible. So thanks so much for sharing with us today your view of, you know, how material science could develop and innovate. It sounds like a really key moment in time for the whole industry and such innovation you're bringing. So thanks for coming on to chat with us.
Jonathan Godwin: Yeah, it's a pleasure. Thanks so much for having me.
Daniel Darling: What an incredible and timely conversation with Jonathan. The work he's doing at Orbital points to how AI models are revolutionizing the scientific method. What stands out to me is how AI is being used to further advance AI itself. And what I mean by that is how Orbital is revolutionizing the data center, the brain of AI, which in turn helps to usher in the next generation of compute and more advanced models. This creates a perpetual cycle of progress and advancements. It's incredible to see decade long research projects being toppled in just minutes thanks to the power of these models and points to a golden age of science ahead. To follow Jonathan and the work he's doing at Orbital, head over to their account on X, @OrbMaterials. I hope you enjoyed today's episode and please subscribe to the podcast to listen to more coming down the pipe. Until next time. Thanks for listening and have a great rest of your day.
