The AI Agent That Compressed 8 Years of R&D Into 2 Weeks

0:00An AI agent can process several tens of thousands of papers a day and then has perfect memory of all the content. So that can be reduced from a month to on the order of minutes. Instead of human scientists, you have what's called A-Lab, autonomous lab. It's basically a high throughput robot that will do 5 ,000 formulations in one morning. When you give that to an AI model, it will give you about 1 ,000 parameters. We can't really interpret them. It's like a different language not meant for us human species to understand, but it works. Welcome, humans, to the Neuron AI podcast. I'm your host, Corey Knowles, and I'm joined as always by the undefeated champion of one more thing, Grant Harvey.

0:46How are you today, Grant?

0:47Dr. Qichao Hu:I'm good. I'm good. It's a little rainy out here in Southern California, which is uncommon. So if you hear the pitter-patter of rain, that's what's going on. Then today I win the rent lottery, I'd like to say. It's beautiful here. Wow, that's rare. I know, right? Right? Well, we'll be joined here in a moment by Dr. Chi Chow Hu, founder, chairman, and CEO of SESAI, a company working on lithium metal batteries and a more transparent EV battery supply chain. With joint development agreements in place already with General Motors, Honda, Hyundai, and maybe more. We'll find out. Now, if you're wondering why this matters for AI, SES AI actually uses AI agents to discover new battery materials.

1:31Dr. Qichao Hu:Their platform, Molecular Universe, compresses years of material research into minutes. And they also use AI on the manufacturing side to catch defects and predict battery health. It's a great example of AI solving a hard physical world problem, not just a digital one. But first, please take a second to like and subscribe to the channel so we can keep bringing you the most interesting people in tech and AI. And with that, Dr. Hu, welcome to the Neuron. Thank you both for having me. It's great to have you here. We're really excited about it. And I guess for those who haven't thought deeply about batteries in years, because I assume the average person probably doesn't, but there's a lot going on.

2:06What problem are you trying to solve at SES? A lot. I mean, I think if you look at batteries, it's everywhere, but then it's a simple device, but then it's quite often the simple device that's actually most complicated, it, especially if actually trying to change it. So I would say 10 years ago, the problem that we tried to solve was a better type of battery, a new material for the battery. And then that's evolved to trying to come up with a new way to come up with new materials. Wow, that's really interesting.

2:44Dr. Qichao Hu:And you're using AI as part of that process. Like we just talked about very briefly, two of the ways that you're doing that. One of them is molecular universe. And perhaps we could talk a bit more about that. And then the other one is Avatar as well. Yeah, yeah. Yeah, so if you look at just battery applications, some applications you need to have higher energy density, basically make the batteries lighter. And then in some applications, you need to make it cheaper. In some, you have to make it last longer. And then each one takes about 10 years. So if you follow the traditional path, and then it will basically take you about a decade to solve each of these battery materials problems.

3:22And that's not a very... Why does it take so long? So there's a couple of things in the battery, and it's similar in life science, in drug discovery. When you have a new material discovery, you go through several phases, right? Basically, first you go through this idea creation phase. Like you have to have an idea. You come up with an idea for this new type of materials. And the second is idea filtering stage. You have this idea, and then you have lots of candidate materials, and you have to filter. This could be millions and then billions down to hundreds. And then third is validation. So you're down to a couple hundred, but then you have to test this.

4:11And then in drug discovery, you go through trials, clinical phase one, phase two, and then approval. And then for battery, depending on the application, you have to do room temperature cycling, low temperature cycling, high temperature cycling. And in some applications, for example, EV, you buy a car and that battery needs to last at least eight years. A lot of times the warranty is for 15 years. So that means you have to test the battery for at least eight years. And there's no good way to accelerate that testing. testing process just takes a long time.

4:49Dr. Qichao Hu:Yeah, that makes sense. And then how does a molecular universe and or avatar help with that process, if at all? Or do they solve a different problem? So again, it's still the same three phases, idea creation, candidate filtering, and then validation. So for each one, idea creation in the traditional human process, that takes about, on average, a month to come up with a really good idea. And also the horizon of a human scientist is limited. On average, a human scientist reads three to five papers because human scientists also have to eat, have to drink coffee, and then get sick and tired. So inefficient.

5:33I know. And then an AI agent can process several tens of thousands of papers a day and then has perfect memory of all the content. So that can be reduced from a month to on the order of minutes. And then second, the filtering stage, that in the traditional process, you have senior scientists, principal scientists coming out with this idea. And then those candidates are sent to the junior scientists basically in the lab to make these things. And then a junior scientist, for example, can try different, maybe 10 to 20 different formulations a day by hand. So now as part of AI, you have this dry lab, white lab.

6:19So the idea creation, think of that as dry lab, basically computer ideas. And then this idea of filtering, this is a white lab. So you go to the lab, and then instead of human scientists, you have what's called a lab, autonomous lab. It's basically a high-throughput robot that will do 5 ,000 formulations in one morning. Lifetimes of work for one human. Yeah. Yeah. And then it's like perfect accuracy, no error. So that can reduce the filtering from, again, several weeks, another several weeks, even months, to just days. And then the last one, probably the biggest bang for the buck, is the validation.

7:02Validation takes a long time. It takes several years traditionally. So once you have enough data and you can train these machine learning models, you only need to capture just the first probably two weeks of testing. And then you will know. You will know it's end of life. So basically take each phase and then you can shrink what was originally years now to weeks, if not days. So this does away with the whole idea of having to actually test a battery for eight physical years. It's able to do this in like a controlled, analyzed setting. Yes, yes, yeah. That's amazing. That's amazing. We've seen this technique being developed and deployed in life science a lot.

7:52Of course, so in R &D, when you develop a new material, you have to go through lots of trials and errors. Say you go through 100 times that don't work, and then that 101st time works. So that first 100 times, you can really use this process to accelerate. And then that final trial, the last trial that actually gets you the breakthrough, that one, of course, you can still take the time, do the full testing, do all that. But just the process before that can be much faster. I guess that explains why batteries essentially went the better part of a century with pretty minimal advancements. Isn't that?

8:38Yeah. Yeah. Almost no. So just put things in context. the materials that are used in batteries, a lot of times these are small molecules, small organic molecules. Then in the universe, there's about 10 to the 60th, six zero possible small molecules. And then since the 1990s, the last almost 40 years, the battery industry only screened about 10 to the third different small molecules. So we explored 10 to the third out of 10 to the 60th possibilities. A lot of room unexplored. Exactly, yeah. That's amazing.

9:20Dr. Qichao Hu:So did you, when you decided to move on lithium metal, was that because you already had a hypothesis that that was just better? Was that just where the industry was moving? And where did you think of bringing AI into the equation to help speed this up? When did that pick up? Yeah, so we knew lithium metal was going to be the best because if you look at the periodic table, lithium metal is number three. Then lithium is the lightest metal we have on the periodic table. And then even lighter than that, you have helium and hydrogen. So in terms of portable energy storage, you can't get better than lithium because it's already the lightest metal we have.

10:00So that's why we focus on that. Now, to make lithium metal safe and stable, for a long time for different applications, it's really difficult to come up with a material called electrolyte for the lithium metal. So you had to come up with an electrolyte material that was stable and safe on lithium metal. So then a lot of the work was coming up with a cocktail, a formulation with new small molecules for lithium metal that will make it safe and stable.

10:31Dr. Qichao Hu:Got it. Got it. And then just to follow up on the AI side of that, did you use AI as part of that process, or did the AI come after you already had the initial cocktail and you wanted to do all the testing and all of that? Almost in parallel. So we were really frustrated with how slow and how much work was taken to test different cocktails for the lithium metal electrolyte. So this tool, Molecular Universe, really came out of that. So we wanted to, okay, say if we have a database, if someone just had mapped all the small molecules that could be used for this battery and have almost like a dictionary there for us, and then we could just go to the dictionary, find the molecules, and then also instead of testing it, then if we just had a model and then we only give it, for example, just the first phase or so, and then we can almost predict the end of life.

11:29And then instead of having a human doing the testing in a very inefficient way, if we just had a high-throughput robot. So we had all these ideas, but then no one really supplied these tools to us. So we built these tools. That's awesome.

11:44Dr. Qichao Hu:Yeah, I love that. Like, does it exist? Well, shoot, guess we got to make it. Yeah. Yeah. Yeah. From a practical perspective, like to someone who's maybe, you know, not as well versed in this, what's the practical advantage of, say, lithium metal over mainstream lithium ion that we use every day in our phones and such? So it's lighter and then smaller. What that means, for example, if you were to put that in a pickup truck, the range of a pickup truck is dependent on the amount of battery power and also the weight of the truck. So if you can make the battery lighter, then that truck can go farther, or you can put one more passenger.

12:29You can put more payloads. And then same thing with drones. so it's able to fly farther, or you can add more passengers and then more payloads. Wow. Okay. So that's really interesting. It makes a lot of sense. I guess up to now, this has really been a balancing act of how much battery can we put in and not cross this threshold where we're losing range based on the weight of the battery. Yes. Right. So if you keep the technology the same and then you're limited to the same power to weight ratio. It's almost like a rocket. You cannot just add more battery and then expect that to go farther because then the more battery you add, also you're adding more weight.

13:14So it's not going to work. So you have to really use a different chemistry. That's when you start looking at how can I make my seats lighter? How can I build a lighter dash? What can I make my motors out of that will be a little light? I guess everywhere you can trim an ounce. This is why like the early days of EV in the 90s and 2000s, there were these tiny cars, right? Like no one liked those cars because they were too small. But then the EVs after the 2010s, they started becoming more practical, like families and then more practical. Yeah. And it's something you'd be more, I like to say that, you know, Tesla kind of brought the cool factor to EVs.

13:56Like that was a thing that was really missing. was the idea that a car's always kind of been an extension of yourself in a way and representative of who you are. And the little bitty smart cars were not very representative of most people in a lot of ways, I think.

14:16Dr. Qichao Hu:It was a hang-up for me. Yeah. Well, that makes me wonder about the current hang-up today, right? Because the EV market today is really interesting, where if you look at it in the US, there's really not an affordable EV that most people can drive. And I feel like that's probably the biggest constraint on adoption right now is like, they're just too expensive for the average person. I'm wondering, feel free to address that or correct that, but what would you say is the biggest technical constraint right now? Is it safety? Is it cycle life, manufacturability, or actually just cost of the materials?

14:50So actually, if you look at EVs today, I mean, the U.S. market is a bit unique. If you look at Europe, if you look at Asia, the EV market is actually quite different. For example, if you go to Norway, if you go to China, basically more than a third of all the new cars being sold are EVs. And then the cost has come down a lot. So quite affordable compared to a regular car. And then a lot of these new EVs are being made by a new wave of car companies. And they really stress the interior design. There's TV inside. There's a massage chair inside.

15:35And also, you can have a hot pot inside this. So the utility is way more enjoyable driving an EV than a regular car. because of all the new utilities. I would say the U.S. EV market is unique in the sense the subsidy went away as of September last year. Now it's... So without the subsidy, the economics changes.

16:04Dr. Qichao Hu:And also U.S. market is a molted market in the sense a lot of cars from Asia, it's hard for those to come in without a tariff. So that also changes the market. So that's not even a battery science constraint. That's just a – It's not. Yeah. Business, political, economic, everything else constraint basically. Yeah. Does the supply have an effect? I know that traditionally some of these metals are very geographically located and acquiring them is difficult, comes with a lot of struggles as well. Does that play a role in pushing these things forward? Or the decision to stay with lithium, I should say?

16:47I mean, it does to a certain extent, but now the supply chain is quite diverse. A lot of the lithium comes from South America, Australia. They get refined in China and also some in Canada, and then they get assembled into batteries. So in terms of availability of this, it's not an issue anymore. Of course, sometimes the raw material price fluctuates and that influences the cost of battery. But in terms of availability, it's not a limitation. Okay. That's good. I was just curious because I know I was thinking of like with cobalt, there were struggles as people were looking in those directions and others.

17:27And I wasn't sure about specifically how the supply of lithium looked. So thank you. Yeah. So cobalt is used in the phone. So, for example, in the foam, the cathode is called lithium cobalt oxide. It's basically all cobalt. But then in the EV, there are two types. There's nickel cobalt manganese, where cobalt is less than 10%. And then the other type is lithium iron phosphate. So it's actually cobalt free. There's no cobalt in that type. So it's not a constraint anymore.

18:02Dr. Qichao Hu:That's good to know. And lithium was a constraint, but it seems like a lot of emphasis went towards making more net new mines for lithium and trying to make it very accessible over the last couple of years. Yeah, and also recycling. So, for example, now we have lithium coming out of mines. So, one example, you take lithium coming out of mines in Chile, and then that gets shipped to China to get processed and assembled into a battery and then sold into a vehicle in the U.S. And then this battery in the vehicle gets recycled in the U.S. And then that lithium, that nickel, that manganese get used for the new battery.

18:41Dr. Qichao Hu:Yeah. So the recycling actually allows you to not go back to the mine anymore. Yeah, that's awesome. And that was like a much needed aspect of the supply chain that I feel like has gotten ironed out recently. Yeah, which was great. We have one not far from here. We had a lithium battery recycling facility that was dealing in like old EV batteries, essentially, and stripping those and preparing the materials to go back. Is there a limit to how many times that can be reused or does it stay? You know, essentially you're looking at. I mean, there's some loss, some loss, less than 10%. So each time you lose some, but it's to the most part, it's pretty efficient.

19:21Dr. Qichao Hu:Wow, that's amazing. And you said like the battery should last at least eight years, right? Like, so in theory, that's like 80 years worth of potentially 70, 80 years. Yeah, exactly. Yeah. Yeah. That's cool. I want to go back to Molecular Universe for a second because you have this great database of all of this chemistry information. I guess my question is, what would be the next thing that you would want to do with that database? Like, are you just making lithium metal as efficient as possible? Are you coming up with new compounds? Are you exploring other material batteries? Like, what can you do with this now that you have it?

19:54Yeah, so it's almost like the Britannica, the encyclopedia, right? And we put a lot of emphasis on the data, and then two kinds of data. Again, dry data and the wet data. So we really want to map the entire universe of materials and all the properties. Not just batteries, but then like pesticides, detergents, cosmetics, oil and gas, paint, basically everything. These materials boil down to small molecules, and then there is not an encyclopedia of all these materials. So our goal is to gradually build this database of all the materials and then map it. So map it meaning dry data and the wet data.

20:35Dry data just use computing horsepower and then compute all the properties. And the wet data is basically we have these high-throughput robots that actually run these experiments 24-7 and then collect the wet data. And so we use the wet data to calibrate the computer dry data. And then we end up with this modern-day encyclopedia. And then this, we can feed it into new models that we are developing for the different applications. So this goes way beyond just batteries. And the goal is to apply this to almost any material R &D. So the other day we had this, so one of our employees is working on a project with a home goods product.

21:22It's basically detergent and he's testing different cocktails for detergent. Yeah. And that's quite, quite similar. That's cool.

21:31Dr. Qichao Hu:I was not expecting that. Yeah, me neither. That is really impressive. And I want to call out something that you mentioned. So you're using AI at the beginning to do your dry work, but you're also using high-powered robots at the tail end of that to handle the wet experiments as well, right? Yes, because if you only do the dry computation, it's not very accurate. I'll give you one example. If you just use models to compute, for example, melting point and boiling point, oxygen molecules, typically you're off by 30, 50 degrees Celsius. But if you have actual data from the wet lab, like actual raw data, and then you use those to calibrate, then that error bar can shrink to maybe plus or minus two or five degrees.

22:20Dr. Qichao Hu:So does that create a feedback loop then where you're using the molecular mapping for the dry data? You're then getting wet data to validate, and then you can feed that wet data back to your map and create a more efficient map or a more accurate map? Yeah. So the dry data really allows you to map a much bigger universe. You can compute, for example, 10 to the 8th, 10 to the 9th pretty quickly. White data, you're talking about 10 to the 4th, 10 to the 5th, so significantly less than the dry data, but that's enough to calibrate the dry data. Okay. So it's like basically, yeah, it's like, would it be equivalent to tuning it, to tuning the dry data?

23:02Dr. Qichao Hu:Yeah, yeah, yeah. Exactly, yeah. And I guess just so I understand, when we're talking about this map, is this a bunch of text data or is there 3D models involved when we're dealing with chemistry? Is it a mix of both? What does it actually look like conceptually? What the data look like? Yeah, what does a map of the data actually look like? Are you dealing with 3D simulation models like that? Or is this all just a bunch of text of chemical compound combinations? What kind of consists of it, I guess, is what I'm wondering. What does it consist of? Okay, so the molecule database consists of just molecule structures, and the structures are in 3D, but then you can represent the 3D in what's called smile strings.

23:44For example, water is H2O, and then you just write O. So you can represent a 3D structure with a string of letters, C, H, O, those letters. And then what we compute and what we measure are these properties. The properties are just in these numbers. For example, melting point, boiling point, energy levels, viscosity, just numbers. So at the end, you end up with an Excel table of 10 to the 9th, 10 to the 11th, eventually 10 to the 60th smiles streams. And then all the numbers, all the properties. That's awesome. Do you have a rough idea of how many of those you're running in, you know, I don't know, a week, a month, a year?

24:32What's – like I feel like there are so many applications for this, like you mentioned, that it goes so far on DVs. What's your compute budget? This is important research for decades. Yeah, yeah. Not enough. Not enough. So for now, the basic dry data we're computing is using a technique called density function theory. That one, if we use machine learning accelerated density function theory, we do about 9 million molecules a day, just the single molecule level. And then once you get to the cocktail level, so that's where you mix three or five different molecules together. Right now we can do about 2 ,000 a day, but we need to do way more.

25:18I assumed it was going to be just a tiny fraction of what your dry is. Yeah, exactly. So how do you choose which from your dry work is going to go and actually be tested? Are there like, we need to test one from this area or is it random? Is it random?