‘Without AI, we're stuck in first, and we need to be driving a Ferrari top speed.’
Artificial intelligence is quickly accelerating drug discovery, healthcare services, product design, and manufacturing efficiency. Now it's here for materials development – and it could be one of the most influential uses of AI in energy.
A decade ago, Greg Mulholland started playing around with machine learning to accelerate product development of materials inside LEDs. After seeing its early potential — and anticipating the rapid evolution of AI — he co-founded Citrine Informatics.
Citrine built a model that helps researchers advance materials for use in solar cells, batteries, electric cars, and sustainable buildings, among other industries. Machine learning can offer researchers new insights on the relationship between processing and material structure, and offer new suggestions for experiments.
For example, one battery company recently created a new quenching gel for fire prevention in a matter of weeks. “A scientist is great. AI is incredibly powerful. A scientist cooperating with AI is exponentially better than either of those alone,” said Mulholland.
In this episode of The Carbon Copy, Mulholland describes the many ways that artificial intelligence is pushing the performance of clean energy technologies in the lab and in the field — and helping clean up the materials that make up the world around us.
Stephen Lacey: Greg Mulholland thought he was going to become an electrical engineer.
Greg Mulholland: I wanted to design new processors and make new faster computers.
Stephen Lacey: But after four years getting his degree in the field, he realized two things. One, that energy and sustainability were a lot more important to him than just building faster computers. And two, that many of the challenges facing power engineering were actually materials problems.
Greg Mulholland: It was how do we get the semiconductors to operate faster at higher voltages and with all these other characteristics that ended up being more about the materials we're making things with the quality of those materials and the specific capabilities that they enabled. I really developed a passion for the materials side of things and started to pursue that, and it has become my life's work.
Stephen Lacey: Greg became a material scientist. He wrote his thesis on metal oxide nanowire arrays for hybrid organic-inorganic solar cells, essentially using semiconducting particles for increasing absorption of the solar spectrum. Then his love of materials converged with another up-and-coming industry. Do you remember the first time that you played with artificial intelligence?
Greg Mulholland: I do. When I was right out of grad school, I went to a small semiconductor company. It was making gallium nitride, which is the material used for blue and green LEDs. We were generating a ton of data about our reactors, about the materials we were creating. I was pretty unimpressed at how we were analyzing that data.
Stephen Lacey: It was all very ad hoc. That's when Greg started playing around with machine-learning tools as a way to accelerate product development. This was around the turn of the last decade and the models were far less powerful than they are today.
Greg Mulholland: It was an awesome success, but it was when I got to business school and met my co-founder who had done his PhD specifically in machine learning for materials, that I really started to understand what the power was. He and I started a project to play around with materials data and see if we could make something of it. In the summer between our two years of business school, we actually invented a new class of thermoelectric materials using some combination of machine learning, artificial intelligence, depending on how you define these things. That, for me, was just a really huge moment in both my career, but also in understanding how these tools can be used in industries that people generally don't think of when they think of digitization.
Stephen Lacey: Artificial intelligence is quickly accelerating, drug discovery, healthcare services, product design, manufacturing efficiency, and now it's here for materials development, which could be one of the most influential uses of AI and energy.
Greg Mulholland: What it allows us to do is not just get to answers faster, but move through generations of materials discovery so, so much faster that we're not just incrementally improving things, but we're actually driving forward the whole industry and raising the bar on ourselves in a pretty exciting way. Without AI, I think we're stuck in first. We need to be driving a Ferrari top speed.
Stephen Lacey: This is The Carbon Copy. I'm Stephen Lacey. This week, Greg Mulholland, the CEO of Citrine Informatics, describes the many ways that artificial intelligence is pushing the performance of clean energy and climate technologies and helping clean up the materials that make the world around us.
At a very basic level, what is the materials industry? It's this vast industry. How big is it? When we say the materials industry, what are we talking about?
Greg Mulholland: It's all of the stuff that we make other things out of. That could be building materials. What are the panels you're using in your house to hang the walls, drywall and roofing materials and these kinds of things. It could be semiconductors, like where I came from. It could be batteries. Batteries are just stacks of materials arranged in clever ways, solar panels. But it's also the titanium that the new iPhone is made out of and the vehicle bodies of cars. It's really everything that makes up everything that you buy.
From a size and scale standpoint, there are very few, if any, industries that are bigger than materials. Just building materials alone is about a $1.2 trillion industry. The whole US economy is about 23 trillion. You can see that when you start adding up all of the input materials to making those end products that we use, it's a really huge number. That doesn't even start to touch things like toothpaste and shampoo, which are really just formulated materials that you're using to an end.
Stephen Lacey: How often are we discovering and creating new materials?
Greg Mulholland: Discovering new materials, what that is depends on who you ask because say your favorite laundry detergent company, it's a chemical, or Apple comes up with a new metal, they're going to make the new iPhone out of some metal alloy that no one has ever made before. Well, in all likelihood, they're using something that we know pretty well. The newest iPhone was titanium. It is titanium. They use Ti64. It's aluminum and titanium mixed together in that ratio. You have this new material. That's an edit on an old one.
Fundamentally, new materials are pretty rare and usually in really advanced disciplines, solar, for example, batteries, but new fundamentally new materials often win Nobel Prizes. If you think about the cadence by which we invent something totally new, it's pretty infrequent. It's measured definitely in years and sometimes in decades. With that said, every single day there are scientists all over the globe trying to come up with the next generation of a material that maybe they know pretty well, but it's a new enhanced version, a more recyclable plastic or a lighter weight alloy. That work is just as important if very much more frequent than the fundamental new discoveries.
Stephen Lacey: I want to talk about the importance of materials in the context of the energy transition. What are some examples of important materials innovations that have pushed clean energy forward?
Greg Mulholland: Materials are the fundamental technology that power the clean energy revolution. They are the things that are going to help us decarbonize our impact on this planet. There are a few particular cases that I think are worthy of highlight.
On the generation side, there's things like solar cells. Solar cells are just simple, layered materials that interact with one another in a particular way to generate electrons. By the way, some of those innovations, perovskite solar, dye-sensitized solar, even I mean basic silicon solar cells have been refined with new materials technologies over time.
You then look at things like wind. Well, wind generation is you've got magnet technology that's really critical, also critical for the EV revolution. You've got light weighting. You don't want your turbines to be heavy. That's wasting good energy. You need to make sure your materials are anti-corrosion because a lot of times these things are at sea in the saltwater and they need to be something that can be robust to that sort of environment.
But maybe most excitingly, and the area of hottest debate right now, is batteries. Lithium ion has long been seen as the winner in the battery wars, but lithium ion doesn't do full justice to the diversity of materials that are in our batteries.
In China, there's a big push for lithium iron phosphate. In the US, we have standard lithium ion batteries that have a lot of nickel, manganese, and cobalt in the materials. Those things are hard to source and rising in costs and sometimes sourced from places where if we took a good hard look at the mining practices, we probably wouldn't be happy with from a person standpoint, a humanity standpoint, and from an environmental standpoint. Environmental needs are only going to be addressed by being able to adapt to the input materials we have and turn those input materials into really highly tailored output materials that allow us to power ourselves all the way around.
Stephen Lacey: This process is iterative. It's rare to find some new material that hasn't been discovered yet. But I'm really interested in how this process works. If I went into one of these research labs, who's involved? How much time is it taking? What are the costs associated with iterating on these materials?
Greg Mulholland: The people developing new materials are incredibly smart, incredibly highly trained people. These folks usually identify as material scientists, sometimes as chemical engineers, chemists. In some cases, biologists actually. There's a lot of cool biology in materials these days. But what they are typically doing is one of two things.
The simplest version is someone comes to them and says, "Hey, I have a customer who really wants a better solar cell or a better, a slightly tweaked version of an alloy or a ceramic or a battery." They take the lessons they've learned from college and grad school and in their professional careers and they say, "Well, if you add a little more lithium to the thing or you sprinkle in a little silicon, you will get a better performance." It's funny, there are times where I've heard very serious scientists just refer to that as pixie dust. How do you magically sprinkle on something that will enhance your performance a little bit? In some cases, that's very doable, but usually that's doable for a while, not forever.
The other thing that people are constantly researching is breakthrough materials. How do I make the disruptive battery or the disruptive solar cell or whatever the thing might be? Both of these approaches have pretty similar risks. First of all, you're relying on just the intuition of one or maybe a handful of very expert scientists, but we all know humans have bias, right? That's good and bad. Bias in the sciences is sometimes just seen as learning, but it's also it can blind you to things that might be possible.
It's also narrowing is the problem. If you are on a team of three trying to crack one of the hardest problems known to humanity where if you do it well, you win a Nobel Prize, I don't want to bet on that. It's a very long process and trial and error-driven process. When we think about materials development, typically new materials, like truly new materials are developed in the five to 10-year timescale. Then they take a long time to roll out because you have to prove that they work and work for long periods of time. Then individual refinements on those materials typically take place every six to 12 to 18 months, depending on the industry and how difficult materials to make, but the costs are enormous.
One of the things that we don't have in the materials industry is huge volumes of data because everything costs money. There is no free experiment. As we as an industry grapple with a change in innovation rates and trying to improve the way we deliver products, there's a really, really big gap between the practices of yore, the practices that led to the Bronze Age and the Iron Age and the ages named by materials that happened every couple thousand years to now saying, "Hey, every two to five years I need something fundamentally different," it requires a fundamentally different way of approaching things than just a combination of serendipity and local expertise.
Stephen Lacey: Into this environment steps artificial intelligence. Suddenly everyone is focused on the use of AI for drug discovery, for directing Hollywood-style movies, for sorting through legal documents and writing legal briefs. Materials discovery is a very interesting application, one that you have actually been working on for a decade. When did you suddenly realize that machine learning could actually be really helpful in discovering new materials and helping with these complex engineering challenges and decisions?
Greg Mulholland: It was in 2013 that we realized that AI could have such a strong effect. I don't think we, even at that time, even in our wildest dreams, realized what effect it would have. Look, I think over the last year and a half, we've seen AI just rise, like you said, in every domain.
But in science, AI is a different question. The reason is that when you go into Hollywood scripts, for every movie that gets made, I have no idea how many they reject, but it's a lot. Then you look at what OpenAI is doing and they're basically scraping the whole internet with some caveats, but they're reading everything. They're doing it basically for free. You can learn from everything that's ever been said, whether it's somebody being crass on Twitter/X or somebody being funny on a bulletin board somewhere or a Facebook post to your family, they're reading all of this stuff and they're able to learn what voices sound like. That's really the core of what ChatGPT and those things are doing.
In science, we have a totally different way of working. The narrative we build is very much around the scientific peer review process. We have journal articles that come out and people write a whole narrative. Every solar paper starts with something to the effect of 10,000 times the amount of energy humanity ever needs hits the surface of the earth every two hours or so. They justify the work and they talk about the work, they present a few examples of why the work was good, and then they conclude. That's the paper.
Unfortunately, from a scientific standpoint, the most valuable thing to AI is the two examples. But you know that those scientists did not just do two experiments. You know that those scientists did months and months and months of work and refinement and failed dozens of times to get a result that they thought they could only to succeed later. There's learning there in those what are known as negative results.
Our goal when we started Citrine was to say, look, these companies are doing work all the time and only a very small percentage is benefiting humanity. Could we actually take the research from within a Dow or a DuPont or a BASF or a Alcoa or one of these really huge companies, consolidate it and allow them to learn faster so that they can make better decisions around what are we making, how are we making it, how environmentally sustainable is it, and can we actually continue to grow this business? Because the reality is the regulatory world is changing very, very quickly and not getting lighter.
What we're seeing is that AI is providing the capability to help these companies build an asset base of knowledge and then learn to develop new materials faster and actually identify the materials that can respond to regulation while still delivering the products that you and I want to buy on our shelves.
Stephen Lacey: How exactly is Citrine using AI? How does the model work? It's there to help researchers, not replace them. How does that co-discovery process work? How do you explain the results? Just walk me through exactly how the model works.
Greg Mulholland: It's actually fairly simple, at least superficially fairly simple. You can think of AI, and Citrine's AI in particular, as a really basically perfect lab assistant. They never get tired. They never show bias. They weight all data appropriately. They start a conversation with you about what experiment you should do next and how you should seek to get to the goal that you are trying to get to.
The way that works is first, as with all AI, we start with data. Now Citrine has collected data over time, but the reality is the companies that work in various materials areas are the ones that have the most valuable data. That's the first piece is being able to bring a company's data into our private platform without sharing it with anybody so that the AI can learn from it for their benefit.
But the problem is, as I mentioned before, the problem is that the data volumes are small. Very often a Citrine materials development project starts with 25 or 50, maybe 100 data points. My team is overjoyed when we see 100 data points. It's like an infinity almost of data for us. That is a good starting point. But I think anybody who knows anything about AI knows that 100 is not really an AI learnable volume of data. That's just not what people do. We take advantage of the other advantage that scientists have, which is a career full of knowledge and learning.
I came from the semiconductors world. I can tell you that from my background, the amount of carbon in a gallium nitride crystal means something. It's important. It's important to the conductivity of the material. I don't need to use my precious 100 data points to understand that.
Battery is the same way. There are battery scientists all over the place that really understand how batteries work. What they can do is they can go in and put in the relationship between the processing, the internal structure of the material, the components of that material into our platform. They can actually draw the relationships or describe the relationships verbally, actually. It's pretty cool. That's before the data touches it. You're teaching the system before you ask the system to learn on its own. We then apply the data to the machine-learning system and it builds the relationships the scientists put in and it learns new relationships.
This is actually the perfect metaphor that I love to use with this topic. We train it like we train a material scientist. You spend your undergraduate career learning from textbooks. You're not really looking at data most of the time. You're being taught by experts who know the domain. Then you spend your graduate career looking at data and trying to learn something new. Our machine learning does that. Then it recommends experiments for the scientists to then decide to run. Those experiments are highly likely to hit the goal parameter range. It's usually somewhere between 10 and 20 target properties. The scientists will go test those things. Then what happens is they will collect new data and they've either hit their target or they feed the data back in the system and the system learns alongside the scientist. There's a rational debate that happens between the AI and the scientist to decide where to go next.
What we find is a scientist is great. AI is incredibly powerful. A scientist cooperating with AI is exponentially better than either of those alone.
Stephen Lacey: What are some of the tangible results? Any particularly interesting breakthroughs or aha moments for some of the researchers using the platform? Have any of these come in sustainable materials as well?
Greg Mulholland: I will say the first most exciting type of result, and this gets to my heart and soul as a scientist and engineer, is when we reveal something to someone that they're an expert in, but they didn't understand. We worked with a group doing research in fuel cell materials. These materials, this was truly breakthrough stuff and real experts doing it, and they used our system. It highlighted that the polarizability of the crystal structure that they were using was actually one of the most important things to consider in whether the fuel cell would work and perform at the level that was expected.
The researcher that was leading this effort said, and she is brilliant, one of the top minds in the field, she said, "I don't know that I ever would've thought of that particular characteristic. It makes perfect sense and I'm going to use it going forward, but I don't know that I ever would've thought that." Sometimes it's just about illustrating to someone what the relationships are into the materials so they can think about it in a better way.
More concretely, we have done work with battery companies, for example, where there are oftentimes fire concerns about a particular battery. Obviously, nobody wants their batteries to catch fire. A lot of times they'll include a quenching gel or some other material in there that prevents the battery from catching on fire. This company needed one of those. They brought their data in. It was just a couple of weeks later, they had a perfectly functioning quenching gel in their product that they would not have been able to find otherwise. It would've been a many, many months long overall project.
We also have worked with a great degree of success in toxicity reduction, and I'll get to the energy pieces too, but you hear a lot about microplastics, you hear a lot about PFAS, you hear a lot about these chemicals that were really attractive when they were invented in the sixties and seventies, maybe less so now that we know what they do to the water table and to our bodies. There's a lot of pressure to get those out. We've been able to use AI to substitute safe ingredients for harmful ingredients.
Most excitingly, on the carbon front in particular, is where we've been able to help companies identify breakthrough battery materials, breakthrough solar materials where they actually can go and optimize materials together.
Let me take a battery example. A battery is not one material. A battery is a layered system of several different materials, canonically at least an anode, a cathode, and an electrolyte, often a separator. There's several other components that could be there. But there is no best cathode, there is no best anode, there is no best electrolyte. You need to build the system that they're all compatible with each other to make the best battery.
What AI is able to do is it's able to say, well, we can optimize each of these materials in turn, but we can also bring them together and optimize the materials themselves across the overall battery performance. We can say not just what's the best cathode, but it's what's the best cathode, anode, and electrolyte combination in the form factor that we care about as a company? Because, as you know, batteries come in many form factors. We've actually had customers where they were in some cases, startups or relatively small companies come to us, develop using our platform a totally new battery material, and then this is the heartbreaking part, go off and manufacture that battery and then come back to us later and say, "Okay, now it's time to do the next rev."
We always want to make sure that we're laying the materials foundation for a generational change in a material, not just, I mean sometimes we incrementally improve things, but making sure that we can actually get to a sustainable set of materials innovations over a long period of time.
Stephen Lacey: We saw recently researchers used artificial intelligence to discover a whole new class of antibiotics, which is something that's very rare. It's extremely rare to discover an entirely new material. Do you think that we will in the coming years see artificial intelligence help us to discover entirely new novel materials?
Greg Mulholland: Oh, absolutely. It's happening today. The unfortunate reality, I kind of like it, but it's unfortunate in some ways, is that a breakthrough drug captures a lot more news headlines than a breakthrough material. The reality is the battery on your phone and the solar cell on your house, or anyone's house 10 years ago was much less good than it is today. That is because of breakthroughs in materials and associated technologies.
What is interesting though is to think about the materials and broader chemicals industry as a bit the grandchild of the pharmaceutical industry. The pharmaceutical industry has the benefit of the FDA. They require audit trails on things. They require long data histories of tests. You have to do it digitally these days, at least on a practical basis.
The chemicals industry doesn't have that. It also means that the chemicals and materials industry tends to be maybe five or seven years behind the pharmaceutical industry in its adoption of new technologies. Those new discoveries of antibiotics or drugs to treat cancers or all of these new cool things, they give me a lot of even more confidence than I already had that the future of the materials industry is loaded with innovation and that both from how do you value a company standpoint, but also from the standpoint of what we are able to buy every day in our daily lives.
Stephen Lacey: What other innovations are you seeing in this field in particular, and what does it look like in terms of competition? Microsoft is using AI. They most recently found a new solid state electrolyte for batteries. Google DeepMind is using artificial intelligence for materials, exploration, testing, the stability of materials. There's a ton of activity right now. How much of it is collaborative? How much of it feels competitive? How do you see this landscape starting to shape up?
Greg Mulholland: It's interesting. Both of those examples you mentioned are really exciting. What I will say is the world of machine learning and materials, I mean, there was a period where I could say that I or someone on my team knew literally everybody in the field because it was just a tiny field. That's not true anymore, but we're lucky to know many of the people involved in those research efforts you mentioned.
I don't consider them competitive. When you look at, take the Google example with the I believe they call it Genome Project, which is its materials project at Berkeley coupled with Google at DeepMind or DeepMind at Google, they discovered 380,000 or so new stable materials that were expected to be stable at least. They haven't made them all yet. That's fundamental work.
Stability does not tell you whether the thing is going to hold up in a crash test. It doesn't tell you whether it's going to be safe for a human to use or touch, but it's step number one. It has to be stable for it to be useful at all. I mean, the thing decomposes in front of you, probably not so useful.
What these folks that are doing the more fundamental research are really creating is a vast palette of new approaches in machine learning, a vast palette of new input materials, and frankly, developing important knowledge that we can then use in the more applied way that Citrine works.
When we work with a company, stability is something we will model and can be really important, but usually we know we're in the realm of stability when we're trying to come up with a new battery material or solar cell material. It's less about stability and it's more about efficacy, which is the Microsoft work started to touch on that. Their work was certainly closer to what Citrine does. But the reality is the number of questions in the materials world that we need to answer, that number is so large that Google and Microsoft and every other company on Earth could devote vast resources to solving that problem and there's still plenty of room for all of us to extend our elbows because the opportunity space just is simply so large and vertically integrated. Apple makes its own materials, but so does Alcoa. There's opportunities there to even interconnect those sorts of companies.
Stephen Lacey: Are you going to get us a room temperature superconductor, Greg?
Greg Mulholland: I've gotten this question before, Stephen. I used to say, "If I could fund the company with Nobel Prizes, I would just go fund the company with Nobel Prizes." But what I will say is when that LK-99 claim was made, what, 6 or 12 months ago, I was very early in somewhat publicly saying this is probably not quite there yet. I think that ended up being borne out. That wasn't based on machine learning, but as a scientist, you start to develop a sense of what the data space looks like, and especially if you live in this data science world.
I think there's a room temperature superconductor out there. My prediction is that it will require at least a variant on known physics, if not entirely new physics to make happen just because that's the nature of those massive breakthroughs, especially, I mean, a room temperature superconductor is a guaranteed Nobel Prize nine months later. That's going to be a big one.
Stephen Lacey: What are the consequences if we don't use AI to speed up the process of materials discovery and innovation? Particularly for the energy transition, you outlined a bunch of important examples, I'm wondering what you think the energy transition looks like without improving and speeding up that process?
Greg Mulholland: Candidly, I don't think we get there by a long shot. Even with new materials, it's a heavy lift because there are those cultural and cost and systemic issues we need to overcome. Collective action problems are hard, but if I were to say materials technology is frozen for the next five years with just tiny little incremental changes, which is I don't want to say without AI we're definitely there, but AI certainly helps us accelerate in a substantial way. I think we are a little bit damned without big, big, big advances in materials. It's not just in one area. It's not just in concrete. It's not just in wood or leather or glass or whatever the thing is. It's about addressing our entire industry as a whole, making it cleaner, and then getting it out into the world as fast as possible.
Stephen Lacey: Well, there's a good chance you may have converted a few budding electrical engineers to become material scientists.
Greg Mulholland: Each path is great. Each path is wonderful. They can't go wrong choosing between those two.
Stephen Lacey: Greg Mulholland, CEO of Citrine Informatics, thanks so much.
Greg Mulholland: Thank you, Stephen. It's been a pleasure to be here.
Stephen Lacey: That's it for the show. The Carbon Copy is a production of Latitude Media. If you want all of our coverage, go to Latitudemedia.com and you can subscribe to our newsletter and see all the other business to business stories we are covering along the energy transition.
This show is produced and written by me. Sean Marquand is our technical director, Roy Campanella III mixed the show this week. You can get all our stories and show notes and transcripts for the podcast as well at Latitudemedia.com.
Latitude is supported by Prelude Ventures. Prelude backs visionaries accelerating climate innovation that will reshape the global economy for the betterment of people and planet. You can learn about their portfolio and investment strategy at Preludeventures.com.
Of course, just please spread the word about the show. Help us find new listeners across the industry for people who are just curious about how the energy transition is playing out. You can post about the show on LinkedIn and X and send a link to your friends and colleagues. We will catch you next week. I'm Stephen Lacey, and this is The Carbon Copy.