🔬 The Limits of AI in Science - Why We Need Self-Driving Labs — Joseph Krause, Radical AI

This is the difference between AI for bio and AI for materials. If you look at bio or, or maybe small molecules as, as a more broad category, you look at selfies and smile strings, right? Which has been a big way to have those materials, those molecules in text.

And then you can use that, and that's because you know the elements, and then you know the bonds, and so you know most of the things you need to know. But what about everything I just told you about the alloy?

Supply chain, cost, microstructure, how you're processing additive versus casting. How do you capture that in, in a string? You can't, and this is what's so hard, is there is no one model that can one-shot a new material that ends up in your iPhone or that ends up on Starship.

That's just not the way materials work. And so there is this really tough challenge of how do you capture all this data and try to bring that back and kind of really improve your AI engine to encompass more than just discovery.

Welcome to Leading Space. I'm Brandon.

I'm RJ, and we are in the room with Joseph Krause, CEO of Radical AI. Joseph, you're in a market that's getting crowded really fast. You have Lila, you have Cusp, you have, um, Periodic all developing AI for material something something.

What are you trying to do that's different? And, and how are you gonna beat the heavily capitalized competition?

Guys, great to be here. Thank you guys so much for having me, and I must start with big fan of the show.

Oh, thank you.

I got to commute into New York City every day, and you're one of the top things that's in my rotation.

Appreciate that.

I always love learning, and I'm a material scientist by training, and so the, the aspects that I can learn from your show, awesome. So super excited to be here, especially in person. Thanks for making it work. What makes us different is our deep belief in experimental data, right?

And I think now you're starting to see the industry pay more attention to this, and you see self-driving labs. I talked about concept everywhere from academia to people like Google DeepMind, all the way through to pretty much every competitor that you've named in the space building an SDL.

It was not always that way. Uh, when we started the company two and a half years ago, people thought we were crazy. "That's CapEx intensive. Are you really gonna be able to pull the data? Models aren't really built for that data today," and we can get into why models struggle in material science, particularly inorganic material science specifically.

And so, "Why are you gonna do that?" And I had a deep belief, my co-founders had a deep belief that, well, in materials, the ground truth is the material itself. You have to be able to make it. You have to be able to test it and characterize it, and then you have to really, at one point, be able to see if it can go into a real application if you're gonna have it used in industry.

And that was where our thesis really started from, was you're gonna build this loop, this closed-loop system, what we call a self-driving lab, that can actually run those experiments, capture that data, and feed that information back to your AI scientist so that it can learn and actually predict materials that are relevant to industry.

That is what our whole company is built around, and for the last two and a half years, that's what we focused on building.

Great. So w-why? Why do you believe that versus the pejoratively the think big thoughts and then, and come up with stuff and then try it later?

Yeah, because so much of what makes a material real is in the latter part of the discovery process, and I mean specifically at the characterization and synthesis phases. So, hey, what did we make? And, you know, does it have some cool properties in the lab?

But, but also after that. You know, we work in a field called structural metals or alloys. So much of what dictates the performance of those alloys is actually in processing. How do you manufacture it? What techniques are you using post-processing and manufacturing that push performance or change performance?

And so, yeah, you can generate a new composition, and that's very important to do, and we do do that with AI. But it's everything that comes after that that actually impacts if you have a new discovery, if that new discovery is relevant to the application space you're going for, and then can you actually make it?

And can you scale it? Can it actually go into the application space and be used by an end customer? Those two and three don't get solved with AI today, right? A model can't figure out your way through the qualification pipeline for a new alloy for a jet turbine.

You have to do experiments to do that. And so that, that ground truth there is really important for us to kinda bring back to and understand, okay, we know what we want to make. Are we actually making those things, and do they actually have the properties we care about, and can we actually push them to industry?

Those latter questions are the hardest questions to answer in materials. And, you know, one thing you always hear about, which is true in materials, is long timelines.

Mm-hmm.

We've heard everything from, like, fifteen to thirty years. Pick your favorite number, whatever you're feeling good of this day of the week. But point is, that's true, and the reason why is it's so fragmented in materials today. You know, academia handles discovery.

On some of, like, light-scale testing, you have small companies that will look at it, typically supported from, you know, the Department of War, Department of Energy, or other government programs like NSF. And then you have the late-stage com- or, like, bigger companies that really optimize their current systems today, right?

They're not focused on our tap superconductivity or high entropy alloys or new ceramics. They're focused on, "How do I take my current material system, make it five percent better, ten percent better, and I capture the margin from that?"

And so there's so much fragmentation across this whole industry that the data never gets shared, and the connection from discovery to manufacturing is typically lost in that process. That's the connection that we wanna bring back to material science.

And, you know, that is what we think the true opportunity is for AI and autonomy in materials, is linking those two together in a fully closed-loop system.

So I wanna dig into that. I have a rule of thumb that I often follow when thinking about things that anytime you change orders of magnitude-

Mm-hmm

... in a, in a scaling system, that your problems completely change, right? So the orders of magnitude of the problems that you're talking about are drastically different, right? When I'm discovery, you have N of one or N of some small number, and, you know, manu- on the commercialization side, you have N of millions or whatever, right?

And so, uh, why do you think that You're capable of solving all those intermediate problems

Yeah, this is a good question, and I think I'll use a real practical example to explain how hard it really is. So in our field with, with, with these alloys, one of the important things that determines property is the microstructure.

And so how does the microstructure form in this alloy so that you can see things like strength or ductility or emissivity or pick your other favorite mechanical property. And so at the generation level, so candidate generation, hypothesis generation, you know, you can predict a new composition.

All right? And, and AI is actually quite good at that. Our, we have our, our, all of our hypotheses are, are generated by our AI scientists today. And it-- you'll take that composition, you'll go synthesize it in the lab.

There, step one, something changes. All right? You might have it not be homogenized. There might be dendritic formation on the surface. You might see different phases, or is it gonna be single phase? Those dictate what the properties look like.

And then once you move past that, you actually go to manufacturability, you know, annealing or thermal processing and actually looking at how do you manufacture it. Is it additively manufactured with powders, or are you casting with actually raw metal?

Both wildly different outcomes in the performance that we like to see. So to answer your question, the first step at solving this is capturing that data, and we do that at the discovery and the testing phase today. So we don't do that manufacturing yet, to be clear, but we do that at discovery.

So we do synthesis and characterization. We have a bunch of characterization tools in our lab, SEM, EDS, XRD, XRF, TGA. And then we all-

Real quick. That was a lot of acronyms. Would you like to explain them now? We can also pause this, and we-

Sure

... wanted to talk about the lab later. If it makes sense today, we can just talk about it now.

We can come back to it if it helps.

Okay. Yeah.

And I'm happy to dive into, like, everything that we do in the lab.

So for now, for the listeners, those are just a lot of acronyms you don't need to know.

And they're just a lot of tools that tell you different things about a material in a lab, and, and we can talk about what they do. Uh, and then we do testing of properties at the lab scale. So we'll look at oxidation performance in our lab today, which is really important to see how these alloys perform in oxidative or corrosive environments.

We'll look at mechanical properties, you know, something what's called a tensile test, which gives you these stress strain curves of a material. And then we'll look at microindentation, and here you can pull what's called the Vickers hardness from a material, as well as we kind of built this proxy for ductility.

It's not an exact measurement of ductility. We kind of pull out if the material is ductile from that. So that's everything happening in the discovery side and moving into the testing side. We have not yet crossed into that, okay, now when we go to manufacturability, but we hope to.

And so back to your original question, if you can capture the data at the manufacturing side as well, now you have the whole suite of what we call, like, the lifespan of the material. I see the hypothesis. I see the synthesis.

I see the characterization and what did we make. I see the early properties showing good results, and then I see the manufacturing and what came out of the back end of that. Does it actually make it to end system?

Now I've seen this lifespan of a material. Now I can use that to go pick more materials targeted at direct applications. That is, that's, that's the North Star of what the company wants to go out and do.

Where do we stand with that today then?

I'd say, um, we're, we're really good at the first part, the discovery and that light scale testing I've mentioned. There are some testing mechanisms that we use externally, like with third parties, uh, where there is deep expertise required in the industry itself.

Aerospace is a perfect example. If they do wind tunnel test or torch testing, we don't have those capabilities at Radical today. And, you know, today so far, we haven't needed to own those. We wanna use third party.

There's a heavy tail.

Yeah, exactly right. Really heavy. And then you look at the cost of a wind tunnel, and you're actually like, "Yeah, I don't know if I'm gonna get enough data from that." Um, and then lastly, we haven't touched manufacturing.

We have spoken with people who do manufacturing, and we do know some of the things they care about. So processability is a really good one. You know, can this material be formed, right? If you're gonna cast it, can I actually move it into the shape that I need it to be in?

We do look at that, um, but we look at it at the small scale, not at ten tons. We look at, you know, grams, two hundred grams, five hundred grams of material, so, so not at that larger scale yet.

But that first section is done. That's running today. We've probably made twelve hundred alloys, uh, in the last five or six months. Three hundred of those alloys are new, novel, never before seen in literature, and I'd say probably ten of those alloys have performance that has got us very excited, uh, on where they're gonna be in the industry.

So that's kind of a rough scale of where we are today from a company perspective.

So that brings up a, like, a follow-on question about how much are you just sort of optimizing within a well-understood space and you're picking permutations that are n-novel but versus trying things, experiments that really are pushing the, the, the frontier of science?

Yeah, the latter. We really are making new materials that push the frontier. So a good example, we work in a field called high entropy alloys, and these alloys are really exotic 'cause they have five to seven elements in the system, all about equally atomic, give or take, and they have really exotic properties in extreme environments.

So think super high temperatures, usually north of two thousand, three thousand degrees Celsius. They have very high pressures, think space, coming back from space. And then they have these environments that can be corrosive, like a nuclear reactor where you're feeling neutron bombardment, or oxidative, like if you're flying, you know, in a defense application or in a jet turbine.

And so these alloys are really exciting here, and there has been, you know, for the past fifty years, the same alloys used in all these industries. And the reason why is these long discovery timelines we talked about. And so this is a perfect example where we're working in an industry.

We're not really creating a new industry, of course. Tur-turbines is a big industry.

Yeah.

It's a great one. Uh, it's having a second tailwind right now with everything we've seen. Um, but what we're trying to drive there is new performance that does not yet exist from the materials they have today. And, you know, we stole this term from, from a gentleman named Charles Kuhlman, who's the VP of materials at SpaceX, which is called concurrent engineering.

And it's this idea that I can actually design my materials As I'm designing my product, right? So as I make a new rocket booster or a jet turbine or a missile or, you know, a, a, a solar cell, I'm actually inventing the new materials that meet the property specs for this.

I'm going back and forth as I engineer them to get to the application. We don't do that today, right? The alloys that are in the plane I flew here on, 1950s, 1960s, 1970s, they might be coated with some CMZs from the late 1990s.

Um, so, so there's this really opportunity of, you know, we're tackling industries that exist, have huge markets, but we're bringing novel materials that historically they never would have had the ability to look at and enough throughput and at enough scale.

So going back to the bottlenecks you were talking about before.

Yeah.

All right, so you said that, you know, it takes fifteen to thirty years to get material through, and part of this is because there's a disconnect in research and, you know, productization. How do you validate that this is going to be the thing which will work and you're not going to get killed by another unexpected, uh, bottleneck in the process?

I, I'm coming from the world of drug discovery, where even if, you know, you think you, you have good, good early-stage data, there's all these things which will get you later on, you know, in different phase of clinical trials.

I'd wonder, are there things like this in materials and, like, where do you think that these are gonna get hit in, you know... What will stop this, you know, super cool alloy you just developed from actually making it to market?

Really good question. There are absolutely those in materials, and if we talk about the alloy example I just gave, one of those areas is called qualification. And so qualification is this process that if it's going in manned flight, it's run by the FAA.

There's also a mil spec one for the US military as well, and essentially your alloy has to qualify to be used in aerospace or, or defense applications, and that process is very slow. Typically a ten-year process today. You have to make a number of different i-ingots of material and run these kind of standardized tests on them to prove your material is usable in those systems.

And so there is a bunch of things that have a "gotcha" later, and I think it's actually trying to capture that data, understand where and why those things are happening that can actually impact your discovery loop.

It takes ten years because like for clinical trials, you have a series of incremental phases. Can you operation warp speed this, where you do these all in parallel, or is it really like you need to do these things sequentially?

There are people working on changing doing it sequentially.

Mm-hmm.

So there, there is really good work right now, um, for example, out of DARPA, where they're looking at new ways to do qualification, right? Where they can use additive manufacturing to go layer by layer and actually look at, you know, can, can we look at qualification of a material that way?

I would say that's a new technique that's trying to just completely redo the way we do that-

Mm

... uh, that, that process today.

So it's not regulatory. This is just... It-

It's... This is the challenging part. I mean, some of the stuff is regulatory in the nature that it's a, it's a government body like the FAA-

Mm

... or mil spec who runs that. And, and you, you have to see almost the humane side of this as well. You know, it is easier to develop, maybe easier is not the right word. It is different to develop a new alloy for an iPhone.

I mean, if it bends in your testing, you can get rid of that or move on or, or even in a worst case scenario, you know, recall iPhones. That's, that's obviously a terrible scenario for the company, but you can.

But if it's going in a jet turbine and you're gonna fly it on a seven eighty-seven, there's a serious bar that has to be met there, and for good reason. No one would want that bar to be removed.

I'm flying back to New York tonight. I certainly, I certainly don't want that bar to be removed, to be super clear, make sure that's on the record. Um, but I think the way we go about that process is very dated, and that's what people are trying to attack is do we have to do it that way or is there another way to get the same result but via a different mechanism?

And that's where AI is interesting, but also autonomy is being really, has been really interesting as well. And kind of as you make the process of manufacturing more automated, there is more sensors, therefore more data, therefore more things you can capture and analyze, and therefore a bigger loop that you can build around that.

That has not been deeply, um, extrapolated in the materials manufacturing sense today.

So maybe one difference between, uh, drug discovery and materials is that in drug discovery you have different phases, each of which is designed to basically not kill people. Um, whereas materials there is in some sense no reason other than budget.

You can't just do this all in one go. Successive levels of qualification do not depend on the prior ones aside from just budgetary constraints.

As well as kind of some of the other mat-matrices you have to pay attention to.

Mm-hmm.

This is what makes materials so hard actually. So a perfect example is supply chain. So probably five years ago, maybe a little bit longer, maybe ten years ago, there were not the constraints that we're feeling today from the metals industry or the minerals industry.

Hafnium, ten to fifteen X in price 'cause China owns a majority of the supply chain. Things like refractories-

Sorry

... tantalum, niobium.

What, what is hafnium?

Uh, so hafnium is an element on the periodic table that, that is used in things like C-103, right? It's about ten percent weight or weight percentage of hafnium in C-103, which is a very common aerospace a-and space alloy that's used today.

So now we're starting to see, you know, in conversations requests around, you know, can you remove that, that material or, or I should say that element from that material? And that's a different problem. There you're actually trying to kind of just meet the same performance specs, but you're trying to completely remove hafnium from that equation.

We've worked on that problem specifically, and we have successfully done that. And so this is where you get back to, you know, supply chain is a concern, cost and margin is a concern. Who is paying that and feeling it?

I can tell you the space industry has a lot, has much more tolerance for high cost. Performance is everything. When I'm designing a new heat shield or I'm designing a new cone that goes on, on, uh, the rocket engine, performance is my number one thing I care about.

Cost is not the first thing I care about. It, it's not irrelevant. I don't wanna pay a hundred million dollars for a nose cone, but I still need... It's still not the top thing I'm thinking about. You think about something like a consumer electronic or maybe even like a medical device application which alloys go into, well, now cost is definitely much more sensitive.

You know, there are probably alloys we could put inside smartphones today, but it would just make them- Unbelievably expensive and probably not tolerant to some of the other things we have to put in there. So there are just so many things about a material that make it so much harder.

And, you know, this is one of the reasons why we deeply believe in self-driving labs. Just to kinda come back to this point for a second, this is the difference between AI for bio and AI for materials, in my opinion, from the materials lens.

You know, if you look at bio or, or maybe small molecules as, as a more broad category, um, 'cause you probably could include some organic materials in that. You look at selfies and smile strings, right, which has been a big way to have those materials, those molecules in text, and then you can use that.

And that's because you know the elements, and then you know the bonds, and so you know most of the things you need to know. But what about everything I just told you about the alloy? Supply chain, cost, microstructure, how you're processing additive versus casting.

How do you capture that in, in a string? Uh, you can't. And this is what's so hard is there is no one model that can one-shot a new material that ends up in your iPhone or that ends up on Starship.

That's just not the way materials work. And so there is this really tough challenge of how do you capture all this data and try to bring that back and kind of really improve your AI engine to encompass more than just discovery and, and even certainly more than just composition.

You mentioned this sort of loop-

Mm-hmm

... where you are-- You're really doing two things with automation, right? One is you're collecting data-

Mm

... one is you're running experiments and manu- building stuff.

Yep.

Right?

Correct.

So how iterative is that?

Um-

And how do you... What is, what does it look like at the different steps where humans could be in a loop?

Yeah, and humans are in our loop today-

Mm

... in a very important manner, training actually and teaching what I like to call the scientist about what they know. We call this scientific intuition at the company. And so what that literally looks like is we have a scanning electron microscopy image.

That's an image that kind of just takes a picture of the material, and our scientist will go in and analyze that image, and they'll make comments in our system, "Hey, I see dendritic formation on this image in these locations."

Right? The AI scientist goes and looks at those comments. And so that's one amazing example of human in the loop where we are trying to download a PhD in metallurgy's brain on when you look at this image, what do you see as a PhD scientist?

We need to be able to replicate that as an AI scientist. So that's one way, one way they're in the loop. The second thing, and we can go deep into this if you guys want to, the lab is not easy to automate.

Uh, it is-

We're, we're getting there

It is super hard to automate. A- and from ways that are like hard engineering challenges, and we can talk about those, and in ways that are annoying. Like, you know, the tool vendors just don't have SDKs or, or APIs-

Yeah

... to, to work with. And so, yes, there is engineering things that we can talk about, but even this idea of the tool provider letting you have access to the data via their software layer was not understood two years ago.

But I can tell you a few very big tool vendors were not too excited about self-driving labs two years ago. They were not jumping to give us, even with payment, access to the software and pulling the data. That tone has now changed.

Uh-

Are they trying to own it? Is that why, or?

From my understanding, and I'm not a tool provider, so they might give you a different answer. But from my understanding, you know, a lot of what they sell is the ability to analyze the data coming out of their tool.

One of the things that makes those tool different is how they actually use the software to, like, generate your spectra, and they like to sell on that, and that's a really big thing for them. So if they give you access to the raw data from that and you no longer need their software, why would you buy their tool?

Then they're commoditized.

Yeah. Now, we tell them, "No, no, no, you're way wrong," and we're getting them there. It's, it's, it's a work in progress. And I do think there's been a lot of moments in AI for science. Number one, it's having an incredible moment, which is gonna be so good for the world.

We can talk about that. Number two, you're seeing, I'd say, academic and national buy-in as well as private buy-in. So you see the Genesis Mission, you see the Department of Energy and the national labs moving this way. You see people like Google DeepMind, Microsoft, uh, other places like Meta either building their own lab or running experiments at someone else's lab to get that data back.

And then you have the private companies that are b- forming self-driving labs and that are looking at the automation of scientific equipment. That has really started to kind of push this wave to, oh, now we... You know, we don't really debate that ALabs or self-driving labs are a part of the future anymore.

It's and which part are we gonna, are we gonna play in that? And that's a better helpful conversation to have now 'cause we can get access faster.

So-

Self-driving labs in biology has been notoriously difficult.

Mm.

Like, the, you can automate certain parts of them, but you inevitably have people who are just walking around, moving trays from one section to another, and it doesn't actually end up speeding things up oftentimes. It, it can sometimes even slow things down.

Yeah.

Um, full end-to-end automation for non-research activities-

Mm-hmm

... or are, you know, for manufacturing, we're really good at.

Yeah.

But when the, the process changes, how do you deal with that? And, you know, have you figured out some way of automating the type of problems you're solving?

You know, the self-driving lab-- First, I think it's important to talk about what a self-driving lab is because this goes in and impacts your answer. There's a difference between an automated lab and a self-driving lab, right? An automated lab does experiments for you, automated without humans and at high throughput, and that, that can be very effective.

A self-driving lab runs research campaigns for you, and there's a big difference. And the way I like to describe it is, you know, one of them is like hands-free driving where, yeah, I don't have to touch the steering wheel.

It'll keep me in the lane. It'll keep my speed set. But when a left-hand turn is coming up, um, I have to pay attention. I have to put my turn signal on. I have to turn the car, and I have to know to make a left.

Now compare that to a Waymo, which I love bringing up 'cause every time I'm here I'm go out of my way. I just drive around the block sometimes to live in the future. Um, you don't need to make a left-hand turn.

Actually, you don't even need to know to make a left. You don't care what route it takes you to get there. You get in the car. You can close your eyes if you want. You can, you know, uh, scroll X.

You can work on a research paper, and then you end up at your destination without knowing how you got there. That is the difference between an automated lab, which you are controlling and just using automation to do throughput, and self-driving lab, where it's actually doing this entire process for you.

So in the self-driving lab, there are things that a human scientist does that are actually very hard. Sample manipulation is a perfect example. You know, when we synthesize these alloys, we get these little pucks that come out. They're called buttons in the industry.

And because you're blasting them at 3,000, 4,000 degrees, they get stuck to the tray. How do you get them out? And you can't-- You gotta be careful because you don't wanna, like, mess with the microstructure or chip off part of it.

Now, th- they're strong enough that you're not gonna really do that, but we had to design custom actuators that go on our robotic arms to be able to manipulate them. And that's not-- or that, that does not really have anything to do with, like, the discovery of this new high entropy alloy.

That's just required if you wanna run autonomous alloy science. And so that's one answer, is there are these challenges that humans either don't face or, if we do, they're very intuitive, right? The button's stuck, I take a little chisel, I smack it, I flip it over, and I move on.

I don't even think twice about doing that. Not, not so much. Two, we talked a little bit about is the software, and it's not just about controlling the tools, but it's running the lab, right? How do I track my samples?

How do I know what sample should go in a tool or should not go in a tool? Is there a quality check where if I look at a sample after it comes out of synthesis, I actually wanna kill that experiment?

I don't need to waste time going through XRD and SCM and the other tools in the lab. I wanna just stop that sample, sta- save it, and, and throw it away. How does it know how to do that?

And this is where you start to bring in, you know, all these different vectors from vision, different sensors in the lab, sensors on the tooling themselves, um, to kind of build this what we call-- we call it operating system, uh, that runs the self-driving lab.

And then the third part is automation, and automation includes what I like to say the connection of the lab. So I have one tool that's automated, I have another tool that's automated, you have this operating system that's running them individually.

The robots, how, are how you connect that. The same way a human scientist would come in, you know, look at the results, take the sample out, and go to the next one, our robots do that today. And so those are the three parts that we really see kinda make up this self-driving lab, and I'd say each of them has their own difficulties.

We can walk through them, some being, like I mentioned, the tool provider, some being no actuators, some being like it's just really hard to load and unload XRD. You have to put it in a hard sample mount, and it's weird, and it's awkward geometry, and you need a custom gripper, and just what it is.

Um, but that's, all of that's kind of forms into what a self-driving lab becomes. We're very good at that for alloys today because the tools in our lab are built for alloys. Some tools sh- share, XRD, XRF, SCM.

You know, even tensile testing, although mostly used in the structural metal space, can be used elsewhere. But I would say, like, our oxidation chamber, that's very suited for the specific customer application we're going after. I would say our synthesis mechanism is directly for alloys.

I mean, we custom-build a tool with a third party to do alloy synthesis in high throughput. That, that's built to do alloys. That's not built to do ceramics or polymers or any other material system today.

Is the goal to expand to polymers and ceramics and everything, or is it we're gonna do alloys and, and, and basically get all the way end-to-end manufacturing on alloys and then expand?

Yeah.

Or never expand?

It's both, but on the right timeline. So the first one is vertical integration, and this is very important. You know, when we started the company, and I'm not afraid to admit, we're like, "We're just gonna do seven different labs, seven different material systems across the board."

"We're gonna capture all this data. It's gonna be amazing." And then we started talking to customers-

It's just a matter of time.

Yeah, exactly. Exactly. And that, that will come back to doing that over time. But started talking to customers, and they're like, "How are you gonna do this? Have you thought about this test? What about when you go to scale?

We need to go to 300 pounds." And we are like, "Oh, no, we hadn't thought about that per se." We were, like, more worried about going to polymers and ceramics and everything else. And so why is that important?

Well, because we are a materials company, right? And, and you see the company talk about this a lot. We really believe in inventing new materials that change the future of the world. We think that is the opportunity with AI and autonomy.

I mean, there are so many industries that we all care about that are blocked because of a lack of novel material advancement. Automotive and aerospace, manufacturing, defense, climate, energy, semiconductors, electronics. You name an example-

What's your favorite example of that? Like, what's a, what's some problem-

Oh my gosh

... that could be unlocked by amazing material?

Aerospace and semiconductors immediately.

Well, like, but what specifically?

Okay. All right. In the back end of line integration for semiconductors, there are particular materials that we've been using for a long time called interconnects. The entire industry is doing R&D here. They are a cause of not having great efficiency and very expensive energy bills, uh, on the back end of that.

A new material would potentially completely remove that problem. That's a perfect example.

Is there, like, a theoretical, uh, like, efficiency that you could achieve, and where are we now compared to that?

We have ideas on materials that would be able to solve that problem, but w- but, but we'll release more on that in the next coming months. We have a spec-

Okay.

We have a specific program we're working on, uh, that's directly around that problem. But it's a really exciting one, and there are estimates from the industry on moving past the current material system, what new materials could bring, but there are other challenges.

Are we talking about, like, two X more efficient, five X, ten X, or one point one X, which is oftentimes quite huge-

Yeah.

But yeah.

Yeah, yeah.

Uh-

I think you could in, in the near future with some of the systems that have been recommended today, you would see, like, a two to five X generally, especially when you think about integration. Um, in these materials you need barrier layers, and there's all these interface things that you have to be able to understand.

Um, I think past that you could start to see it push over ten X. I don't know to what level, but there are cool, exciting materials that we are-- we, we'll, we'll share more about in the future.

What's the, uh, validation timescale for this?

In which way? The lab to start working on these materials or a material in a new chip in the iPhone?

So you have a material you just created that you think is going to revolutionize the world. How long do you think it's gonna take for that thing to get into the, an iPhone or a GPU, NVIDIA GPU?

That's still long.

That's still long.

That's still pretty long, yeah. Um, we're, we're new-- we're early in semiconductors, and, uh, actually, what's long about that is that when you start from ground zero, you have to build everything from scratch. And so even the way that we do material testing for that industry today, and I would say they're one of the industries that's actually f-far ahead of everyone else because they invest so much in R&D, and materials really make or break some of their performance.

Even there, you know, that integration timeline is very slow. And number two, no one can get enough chips, and so everything is delayed in that industry, which is important. I even saw today or this week TSMC telling, uh, ASML they're gonna hold off on some of those new tools until they get through their twenty twenty-nine production run or it was something story like that.

That industry is-

Just 'cause they're go, go, go.

That's what I believe, yeah. Um, so I'll, I'll send-- I'll see if I can find it after and send it to you guys. But I was like, "Man, this industry is really getting pushed to their limit." I would say something on the alloy side, like aerospace-

Mm-hmm

... we feel good opportunity, three to five-year timeline. Um-

It's very short.

Yeah, correct. I think it'll be an application not like manned flight. Like, I don't think it'll be a jet turbine because of the, um, constraints there with humans. I think, like, defense and space systems, though, definitely doable in that timeline.

There have been examples in the past, people that have done that in that timeline, and we feel very confident in our ability to try to execute on that.

I wanna get back to the validation question-

Sure

... because I feel like this is the crux of automation.

Mm-hmm.

Right? Is it every AI engineer who uses cloud code or something has the experience of one-shotting something, and then, um, and then, like, w- and it-- when you look at it, you're like, "What is this garbage," right?

Sure.

And so, um, if-- and if you're talking about what is essentially active learning, where you are hypothesizing-

Mm-hmm

... manufacturing, testing-

Mm-hmm. Mm-hmm

... and then forming hypotheses on that and doing that over and over again, if you-- y-your mistakes will obviously compound as you do that.

Yes.

And so how, w- how are you thinking about this?

Yeah. We call this the negative results, these mistakes, and we do have a version of this loop built already today. So it doesn't include manufacturing data like I mentioned. We don't have manufacturability in the lab today, but it does include synthesis, characterization, and those early property tests that I described to you guys.

And what this system does is kinda have this AI scientist that is really good at designing these campaigns I talked to you about. So it can make up these campaigns. It comes up with the number of materials it, it's confident it wants to go make and test, and it'll launch that campaign.

It'll send it to the lab. It'll start running autonomously, and then it will go through the whole characterization suite, and we'll get all the data. That data is all pulled out autonomously. Some of it is analyzed with machine learning models like computer vision.

Some of it, again, has a human in the loop analyzing it as well. And it'll get put in our database where that AI scientist will look back to when it designs the next campaign on the follow-up of that.

So, you know, it's this active learning loop that is campaign by campaign basis. So it is not every experiment. Honestly, we don't, we don't need it to be that fast. We actually wanna take a few shots and get enough data back to change our hypothesis.

Mm-hmm.

But it is very rapid. I would say, you know, we probably could run probably seven to ten different campaigns running right now inside the lab across different systems. Uh, and, you know, you're updating those daily, um, or at least every other day, uh, with the results that you're seeing from the lab when it comes out.

And what does a human do in that part? Because, you know, I find, okay, so doing, you know, um, transcriptomic analysis and whatever, that Claude does some of my work, and then I end up, you know, sort of course correcting a lot.

Mm-hmm.

Is that kind of the gist of what is going on in your lab as well?

In some parts, yes, though some parts are more complex, to be honest. So synthesis is a really good example. We still have PhDs in metallurgy running our synthesis mechanism. Uh, that tool is not yet fully automated.

Mm-hmm.

Uh, it should be automated by the summer. That's the custom tool I was telling you guys about that we're building with that vendor. And it's not just, uh, opening and closing that's automated. So the synthesis mechanism itself requires automation.

So if you've ever seen how you cast these alloys, I mean, pretty much you take a plasma torch, uh, and you, and you blast them, and you melt-- You take these raw precursors, and you melt them down into a liquid, and then you cast it, and it solidifies into the shape that you cast it into.

So there's a lot of intuition in that. Like, a scientist stares at it-

Yeah

... and, and looks at, "That's not melted yet. Let me hit that corner there." So we have models built that can actually start to learn how to do that at the equal performance that a scientist can, a human scientist can.

So we're getting up there. So that's not fully automated yet. Characterization is fully automated. Uh, we just have scientists annotating images after or results after to train the scientist on. All of our characterization tools can be loaded, unloaded, and controlled with our back-end operating system to do characterization.

Property testing, two of the three property testings are fully automated. So microindentation and oxidation are automated. Tensile is almost automated, should be automated in the future. So we're-- That's where humans are, and that's where humans aren't. Now, after the process, for generation, all of our materials are generated by our AI sci- AI scientist today.

Occasionally, a human scientist will try to compete, and I, I love telling this story. They hate when I tell this story. Uh, and it'll-- They'll throw in a composition, and then the AI scientist say, "Get that out of here.

That's not strong enough." Or we'll see, how do we do, how do we think about something new? Actually-

So you're, like, red teaming.

Yes. That's a good way to describe it. Yeah. Uh, I don't know if they would call it that. I think they would call it losing their job.

No.

But, uh, they're, they're not, actually. They're actually super important to the, for the, to the process. And I think what's cool is sometimes the scientist will recognize a new learning. "Oh, interesting that you threw that element in there."

Yeah.

Um, that's cool. The other part of that is going places where human scientists won't go. And we have this beautiful chart that I can show you that shows in the high entropy alloy space all of the different elements that in publication, so in literature, what we can access, where all the places scientists have gone.

And then we have a second overlay on that chart of where our AI scientist has gone.

Mm-hmm.

And it's moved into elemental families or, or alloy families- No one has ever published on before. And the question is like, why?

Mm.

Like, why did it think to do that? And so we ask our scientists, like, "Why did you never go there? Why did you never use that element or that element?" And its answer is like, "I just didn't think it would be-- I didn't think it would work with the other elements that are in that mixture.

I didn't think it would cast. Um, I thought it would evaporate when we tried to make it, and I, like, it, it didn't. We were able to synthesize it. I didn't think it would work in the microstructure or, or would cause grains to be not what I was looking for and would not get the mechanical properties I thought, so I just never considered it," but it actually works in that formation.

And so there's this really interesting feedback where, you know, now the scientist is getting good at exploring places that I'd say humans have a biased naturally against, even though it might be unknowing bias, and that's, that's the huge power of an AI scientist.

But is part of this just that you have higher throughput and that you are letting the, your AI scientist do its own thing? I wonder what happened if you took those same scientists and say, "All right, no constraints, just go crazy.

You have-- You can do whatever you want to." So the, you know, i- if you have an AI scientist that doesn't have the-- it may not have preconceived notions, I'm actually honestly kind of surprised that it's not just sort of reiterating what is known.

Mm-hmm.

But I, I wonder how much of it is, um, that you've just turned up basically the temperature in your, uh, sampler and-

Yeah

Yeah.

Yeah, that's an important metric.

Yeah.

Processing is really important. So, you know, to be clear, there is times when it, like, goes to what it knows, especially when it pulls in literature. It's like, "Oh, this is where it is." And actually, literature's a great teacher.

Like, if, if things work, it's actually a good place to, to ground on why they work and try and understand why they work. That's a big problem for the materials field that we can talk about. But, but it also has this, you know, good ability to, because it is high throughput, not be afraid to test.

You know, when I was in my PhD, oh, I probably did fifty experiments a year. I don't know, rough, rough estimate-

Mm

... you know, something around like that. So every experiment's kind of important, right? And, and not to mention, like, the mental load. It's like two weeks a time to, like, fabricate this thing and synthesize it and then go test it.

We don't think like-- Or the scientist doesn't think like that. The AI scientist, excuse me. It's like I, I, I'm making eight of them today, twenty of them today, and it-- o- once that tool is done, I'm making a hundred per day.

I don't really care about taking a shot on goal and, and learning from that shot. So it's like a mindset shift there.

How much does one, one experiment cost-ish?

So it depends on what elements you use.

Mm-hmm.

So some elements, platinum, palladium, are much more expensive than aluminum, titanium. Anywhere from, like, sixty bucks up to three hundred bucks.

Okay. And what's your throughput?

It's all element dependent usually.

What's your throughput?

So today it can be anywhere from eight to twenty. That depends on elements as well. Refractories particularly. If we're doing refractories, they are much harder to cast, and so we go down to that eight number. If you're doing things like titanium, aluminum, your standard alloys, Ti-6Al-4V, um, tho- those are much easier to cast.

They, they melt immediately or quickly. Go to a higher throughput to twenty. We should be at a hundred per day regardless of system by, like, the June, July timeframe, rough estimate, um, give or take.

This is across the entire lab and not, like, per, like, workflow.

Yes, that's correct.

Okay.

That's across the entire lab.

Okay. Eight to twenty, and you could conceivably have humans sort of inspect many or most of these, or is that, is that a too-

I don't know. You could. You know, one-- So one, one thing I wanted to touch on, you just reminded me on that answer is AI doesn't operate in the same dimension that humans do. Let me explain what I mean.

When I was a scientist, you go through a very serial-based process. I make a new hypothesis. Actually, backtrack. I read a bunch of papers. I make a new hypothesis. I might run some computational workflows, DFT or MDLnet. I go synthesize it in a lab.

Then I move to my characterization. I study my characterization for two weeks. I get a new idea from an image or something I saw. I circle back. I do that whole process again. That's what a human does, and it's very serial in that if I take a hundred SEM images, I don't memorize all one hundred.

I'd love to think I could. My advisor would have loved if I could, but I, but I, but I couldn't.

Mm-hmm.

And so I pull, like, one thing out of that or a couple things out of that that I wanna learn. Now switch over to the AI scientist and that same process. Now it's parallel. Now I can read a hundred thousand publications and then directly compare them to a hundred thousand SEM images at the same time, in real time, and I can study, I can learn, I can memorize all of the things I'm seeing in those hundred thousand SEM images and draw direct conclusions back to my papers, back to my hypotheses, or back to my, uh, mechanical property testing, right, where I'm seeing what actually comes out.

I can't do that as a human scientist. And so this parallel nature allows it to operate in a way that just human scientists don't, don't have the ability to do.

Okay, but you're still talking, like, order tens or hundreds of, uh, materials that you're producing per week, month.

Yeah.

The overall scale is not that large compared to even, like, b-- a lot of biological methods, you have ways of scaling up to, you know, millions-

Millions, yeah

... or if you're, if you're doing, uh, let's say n- in next generation sequencing-based assays-

Mm-hmm

... you can billions, you know, whatever. So, you know, this is much more remisis- reminiscent of, like, ligand-based modeling where you're, you're really looking at a small number of examples and you're trying to pull out, like, local patterns.

For, you know, small molecules, you can-- you have real predictive power. This is like, these are useful techniques. But the almost universal rule from experience chemomat- chromatics is oftentimes by the time those become useful-

Mm-hmm

... the actual scientist can just go out and do it. Like, they could have designed it by the, you know, they could have found the molecule they were looking for-

Sure

... um, without using the AI model by the time the AI model gets there.

Yeah.

Sorry. So this is, like, a very specific kind of regime, but it's one that has been well established. So I'm kind of wondering, since it seems like the timescales and the sort of data, it seems very reminiscent, it is sort of surprising that this is actually that much more effective.

Yeah, two things there. So number one, throughput is a big important number. So to our knowledge and what is publicly released, so if there's someone that has done it behind closed doors that I'll know about-

Yeah

... please, we'd love to talk to them. Uh- The largest alloys program was the MACH program. It was run by DARPA, MGE Aerospace. They did 500 alloys in about 12 months. They did a bunch of s- um, kind of AI and simulations on the front end of that, and then they synthesized 500 new alloys in that whole year.

So that's kind of the benchmark, I would say, for how many alloys someone can do in a year.

Mm-hmm.

Again, we're, we're trying to do 500 in five business days. We've done 1,200 in three months. So kind of a order of magnitude step up there, uh, that we're moving to. The second piece of this is I think what's really challenging in the alloy space specifically, and I think it's probably specific to alloys, though I do think it will carry over to some other industries, is that there are so many variables that go into, like, determining your end product and then your end properties from that product, and that makes it actually harder to go do discovery on because there is endless amount of potential combinations.

I mean, there are like 10 to the 40 different potential alloys that you could go out and synthesize. So how do you do that? Even if you could do high throughput, to your point, it's still not that much high throughput, right?

We, we think it would take humans seven million years to go make all of them. So what do you do even if you're only doing 30,000 a year, right? Well, the screening mechanisms here are very helpful, as we all know.

That's where AI is great. But I think the other point is actually because the data is missing from the industry, we don't have experimental data, we do see results from 100, 200, 300 experiments quite aggressively. Uh, we see 300 new alloys that we've never seen before from the experimental results of 1,200 alloys that we've run today, and you, you probably think each alloy probably has anywhere from 50 to 150 different data points, depending on how many images you take, how many spectras you run, et cetera.

That's a fairly small data set uh, to get-

Very small

... Thank you. Someone-

Very small

... very small, and it's funny, when I talk to the ML side, they're like, "How are you gonna get to millions of, millions of data points?" And I'm like, "I, I just don't think you need to." We have not seen that you need to to do new discovery yet today.

We have a bunch of new discoveries, many that are going through, you know, patent protection and that we're talking to potential customers about. We just haven't needed millions of data points. And, you know, this brings up, I always get in arguments about compute.

Uh, I, I got in an argument at GTC about this. Like, we don't, we're not compute constrained in the materials industry. Um-

Yeah, you're m- you're making stuff.

Yes.

Yeah.

We're, we're experiment constrained.

Yeah.

And-

I mean, this, this is even what I do, which is computational.

Mm-hmm.

It's m- oftentimes dominated by data movement.

Mm-hmm.

Right? It's not like I can... Uh, I h- I see these people, and I'm jealous of them that they have, like, 14 Claude sessions going, and, like- ... and they, they have all these different experiments going, and I couldn't do that just because I can't move the data around fast enough.

Yeah, yeah.

Yeah.

That's, that's almost a good kinda comp to our world of it's not a model problem. It's not, it's not a language problem. It's not, we don't have the same problems there as an experiment problem. It's really how can you run enough experiments to start to change the output of an AI scientist and capture data you need to discover something new.

So for us, it's really about, that is our bottleneck. That is the throughput. That's why we're so bullish on self-driving labs. That's why when I start the convo, "What do you guys work on?" It's all about the self-driving lab, the autonomy, the experimental data.

I mean, we're trying to build, you know, the protein data bank, uh, for, for, for materials, and it's much more complicated than, than just crystallography structures or, or whatever else was in there. Um, there's all these different properties you talked about today.

It has to be inside that data set to make it relevant, so it's hard. It's hard to do.

So that reminds me of April Kulick's... or sorry, Heather Kulick's episode-

Sure

... where she said that there is no AlphaFold-

Yep

... for materials.

Yep.

So first of all, do you agree? Okay.

I do. Uh, I think that, I think you can have AlphaFold moments for specific areas of materials, like microscopy.

Mm-hmm.

Perfect example. Reading, you know, using a segmentation model on SCM images, our team does that today. Like, that's a cool, whether you call it, like, AlphaFold moment or not, I don't, but there is a real world where models make a huge impact on being able to do that.

What I don't think you can do is go from, like, "I have this new hypothesis," to, "Oh my gosh, I have a new material. It's scaled. It's done. It's in products in your iPhone." You, you can't do that today.

I would agree with that statement.

Okay. But, but even then the AlphaFold solved a scientific problem-

Mm-hmm

... which is how do you take, uh, a protein sequence and figure out what the three-dimensional-

Right

... structure of that protein is.

Right.

And I wanna put in so many caveats so that none of my, uh- ... structural biologists-

Yeah, yeah

... correct me. But anyway, yeah.

Luckily, I'm not a structural biologist-

Yeah

... for everyone watching, so I get a free pass.

Yeah. But, but, like, the thing that seems useful here is that you put in a, uh, chemical formula and a, some sort of processing-

Mm-hmm

... and what you get is a, so it's essentially what you call the microstructure-

Mm-hmm

... or which is something that you can get part of the information out of-

Sure

... X-ray diffraction but not all of. Um, so that problem sounds much harder in a lot of ways than the biological problem.

Agreed.

Can... Okay, yeah. Can you maybe explain a bit more about that?

Yeah, and it, and it's even, um, even harder than what you just described, meaning there are probably things you don't know you should test for yet or that you might see when you go to scale that you did not know that you should have predicted or been paying attention to.

That's really hard to build a data set around. And so kind of what I like to tell people that they do ask about this and, you know, can you build the, the data bank for materials, like, well, I mean, the, again, if you wanna do SEM images, scanning electron microscopy images, and look at the se- you know, use a segmentation model, sh- yes, you can build a really large data set of SEM images that are very good at finding dendrites or cracks or defects in a, in material, and the model will be very, very good at using that to predict, you know, and relate that to mechanical property because you're looking for what crack propagation does to strength.

Okay. Tracked. But that's not the same thing as, okay, can you, can you make a, can you atomize it? Can you make it with a powder? Can you use that in manufacturing?

Mm-hmm.

Can you cast it? Way different thing. Related, obviously, the microstructure relates to that problem, but just because you understand the microstructure, just because you see and can predict crack propagation, does not mean you're necessarily gonna perfectly nail manufacturing.

So man, there are, like, just so many things that we see stack up and some things that we, we learn a lot of new things. Uh, every... We, we think we know everything, and then we go somewhere, and we learn new things along the way.

So I do think it's Multifaceted for sure, and each inorganic material has different constraints, right? I talked about that supply chain. That's very relevant for, and defense applications are a perfect example. Um, supply chain is not one of the things we worry about in consumer electronics per se.

You know, like Ti64 is understood. I mean, it, it's available. Like yeah, the people care about where it comes from, but that's not the same as the critical minerals focus that you see the US have today on where are we getting these minerals that we do not control the vast majority of.

Different problem there. So different inputs that are gonna put in to get an output. I feel like that's why materials are so hard. It's just all of this other data that comes after just discovery. What I always tell people, the second you design a new material, that's a milestone.

The second you synthesize it, milestone. The second you characterize it, milestone. That is not a new discovery. We count a new discovery when you pick up your phone and there's a new material sitting inside of it. That I think is a fair c- uh, claim on new discovery in a scale material.

So as you get past-- So there's fallout in every one of those steps of-

Mm-hmm

... every milestone. And presumably in manufacturing, there's s- separate steps where there's fallout as well, so that as you get closer and closer to the consumer or the application, then you, um, you have less and less data.

Mm-hmm.

Right? So how do you-- How, how fundamentally, how do you get over that?

Yeah.

Because I think in pharma right now, people are starting to think about there are sort of rules of thumb-

Mm-hmm

... I would say-

Mm-hmm

... that, um, can be used to do reverse translation back from the clinic-

Yep

... to the discovery process.

Mm-hmm.

Uh, you know, like how do you do that in material?

You know, it's funny. Uh, I love telling this story. One of my advisors, uh, at the company, he was at 3M for thirty-five years, and we were asking him about manufacturing. Like, "Hey, when you guys go to manufacture, like what you paying attention to?"

And this is early. This is like three months into the company. He's like, "Whoa, you guys got a lot to learn." And I was, "What do you mean? I'm a material scientist." He's like, "Different worlds." Um, uh, so interestingly, one of the challenges, he's like, "You know, the hardest part about the data you're asking for or inquiring about, that is a thirty-five year, uh, trajectory person at the company that knows exactly where to turn the knob on whatever manufacturing tool you're talking about."

Yeah.

And so what you're asking for is, you know, his or her ability to know when to turn the knob right to that spot at this specific moment. How do you capture that? How do I give you that? I mean, even if you assign a formula to it, is it the same every time?

And again, this gets back to intuition, which we've talked a lot about today, just in the manufacturing sense. And this is the hard part. Um, I don't have an answer for you on manufacturing it 'cause we haven't done it.

But I do have a lot of answers on the discovery side, where we've had to look where is intuition important. I talked about the casting of alloys, that, that torch earlier. That's one of them. Reading SEM images. That's one of them.

Looking at XRD spectra and identifying, uh, identifying phases and how strong are the peaks, now that's arbitrary. That's really strong. That's kinda strong. That's not really strong. That's terrible. What do any of those mean? I mean, I can like guess what they mean, but if you look at an XRD, you might not get it perfectly to what, you know, you or you or I think about that.

So this intuition aspect is so important. This is why we still have humans in the loop 'cause you wanna capture that. Now, when you go to manufacturing, we think we'll have to do the same thing, and we think the opportunity is you can actually rebuild those processes fully automated.

And so you can actually put all the s- all the sensors a- and all of the, um, capture mechanisms, for lack of a better phrase, in place, so you can kind of bring all that back. That's a hard problem to solve.

And to be clear, we have not solved that problem yet. We are certainly still at the discovery and the testing side of that, but that's where we wanna get to. How do we get there quickly? Partners. We do talk to a lot of companies in our field who make materials at scale in, in the alloy space who are thinking about this.

And they look at it from a different lens. You know, they're not all hype on AI for science, and actually I'd say a lot of them are, are kinda bearish. They're like, "You don't know what we know, and we've been doing it for a long time."

And that's okay. I think that's healthy. Um, what they do know and what they do bring to the table is actually we, we do have that intuition. We will tell you when you show us a family of elements what we think is gonna work or not.

And we might be wrong, but we can tell you why we think that's gonna happen, and we can tell you why that relates to aspects of a business that are important, like supply chain, like cost, like performance under certain environments that don't exist in others, extreme environments, for example.

Perfect. Temperature, pressure. Those are really important info- That's really important information that you wanna bring back. That's how we get there in the near term un- until we can do it ourself, is you partner with people who wanna bring in th- this discovery, this turbocharged engine, you know, to their process.

So, okay. Right now you're, you're still finding, refining that process in the lab.

Absolutely.

What are some war stories from the lab?

Uh .

What... Give me, give me your best like-

Oh, that's a good question. I love that question. Okay, so we bought the first tools in the lab. Um, oh, I can get in so much trouble for saying this, but it's not-

Say it anyway.

Say it anyway.

Say it anyway.

Um, uh, the, the pleading the Fifth. Um, so some of the tools don't let us interface with the software. We now pay for that software, by the way, and, and like we love that tool vendor. So we were very strategic about how we got access to that.

Uh, and the engineering team, software engineering team was, was smart about how we could do that. So that was a whole two-week sprint that we had to figure out how we could probably try to programmatically, programmatically control this tool.

Um, second-

So what's the juice, man? Come on.

Uh, you, you can look into the things that are running those tools, and you can find out how you can control what you wanna control.

Fair enough. Fair enough.

Uh, so that's one. Um, oh, my comms team is gonna be so mad at me for that one.

We can, we can cut it.

No, no, I'm kidding. I'm kidding. I'm kidding. Um, so I think, you know, one of the other war, war stories that, that we saw early was how interdisciplinary the team needs to be. So we knew it was gonna be interdisciplinary going in.

I think each field we kind of assigned has splintered into even more fields. So, like, you know, materials at large, that's what we knew. You're gonna have computational and experimental, okay. But mechanical engineering, you know, we have real mechanical engineers that build tools, design tools, and put them together.

We have mechatronics engineers that design all our own custom mechatronics to make those tools run autonomously. Obviously, in the field of mech E, but they really have those-

Completely different jobs almost

... exactly, that's right. Software, certainly your standard full stack and, like, building the operating system, but then certainly more on what we call, like, applied ML or kind of like I come from that software background and I'm, I'm applying the systems that we are building, like pulling out images from SCM, i- into the AI scientist, and that was interesting.

Um, robotics, things like path planning and perception, areas we probably didn't think we would need as much as we need today, just because we kinda thought, "Well, we'll use op- what's, what's, uh, open source and off the shelf today," and then we started to realize that what the scientists were doing were very intuition-based, and that's, like, the perfect place where perception or CV or can be really effective.

Um, I do this with the torch. So man, all of these different fields have splintered, and I think we had to build the plane as they fly it. So, so the startup mantra is where we had to continually add people.

Ironically, that has now built a huge moat. You know, in inorganic science, it is not easy to build self-driving labs. And man, if you sent me back two years ago, I'd be like, "Phew, that is a tough path to walk through."

Told you, man.

You know what I mean?

Yeah.

That's a tough path to walk through, and now of course it's a big moat for us, where we're, we're like, "Yeah, it's not about a robot in front of a tool. Go ahead, put a robotic arm in front of a tool and then watch what happens."

Everything else I just talked about will come the second you do that. Now we feel so much farther ahead from the industry on really running self-driving labs for inorganic material science. So that war story, you know, it's funny, we laugh about it, but now we feel a, as actually a huge win and a big edge for us.

Is there something special about this moment which enabled the self-driving lab versus, you know, five years ago, 10 years ago? You've been in the... You've been doing this for a while, so why, why now?

A couple things. One, AI for science is important, right? Like, if I take force fields, machine learned interatomic potentials, they're what, like, I don't know, two or three years old. What- whatever the exact timeline that they're old is, I mean, that's interesting.

You can actually start to really do some parts of your process faster. And although they're still in the computational sense and the simulation sense with things like DFT, they're still very important to that funnel, to move-- to, to, to sharpening that funnel and moving faster at the top of the funnel.

Two, robotics are just better. First of all, they're cheaper. Second of all, you can do more in actuation, custom grippers, different systems that-

Mm

... you know, even we can custom build or we can get. And then three, the buy-in from the tool vendors, like I talked about earlier. Again, I think two years ago we even saw a difference than we see today, where there is a lot more optionality.

You know, I'll give you a perfect example. Some of the tool vendors now have software teams. Or if they had them before, they weren't focused on this problem, they now actually provide support on, "Here's how you can work with the interface," right?

That's a big change actually for the industry. So we've started to see it just become easier to build self-driving labs from an infrastructure and a hardware perspective, I think most importantly been the biggest change. I, I think some of the other things has just been also, like, the, the, the excitement.

I mean, everywhere I go, AI for science is a talked about area. Uh, I was here this week speaking at Unlock. You know, kudos to, to, to Michelle and, and the Medra team. Just unbelievable energy in the room from so many different builders across so many different companies and fields talking about AI for science.

And then I think even, like I mentioned, at the national level, I mean, one of the things that we did early on was spent a lot of time in DC, both on the Hill, at the Office of Science Technology Policy, at the Department of Energy, Department of War, you know, letting them know, "Hey, if you wanna be competitive in science, you guys really need to pay attention here.

AI for science is a serious field. Self-driving labs are a serious field. There are other people who have already built these systems. We really think it is national infrastructure that should be built out." And there's been a big buy-in on that at the, at the federal level, at the state level as well.

So I think you have so many tailwinds that are just, like, pushing this industry forward, plus the excitement, plus the venture dollars coming in that start to solidify some of that, and I think soon we'll start to see some results.

Like, I think we're still waiting on a big result that we see from someone. Hopefully, we're one of the first ones there. I think that'll be a cherry on top to the last tip over, where the customers who are, you know, coming in with restraint or, or caution really see this, "W- we didn't-- we, we can't believe you did that.

We've been working on this problem for X amount of years, we've never had an output like that in that period of time. I'm convinced. I'm a believer. Let's talk about how to work together." And I think you see, look at robotics, I think you're seeing that moment happen already.

Like, it feels like we're on the upswing of that for robotics where, I don't know, two years ago, of course, there was interesting work in robotics if you were a nerd like me and you were, like, reading about it in your free time.

But I, I still feel like when I talk to founders in that area, they're just now getting over the hump where, you know, supply chain and logistics companies and, and the big warehouse companies and even big, uh, humanoid companies are now like, "Oh, okay.

Now there's real foundation models that we wanna pay attention to. We wanna push from the 99% to the 99.9999% in our foundation model technology." I think we're gonna have that for science over the next two to three years.

I do. Once discoveries start coming out and this field continues to mature even more. Remember, we're early in this field. We are a couple years in from an energy, from a community, and from a, from a new technology perspective, so.

So, um, going to the competitive landscape a bit, I mean, we, we talked a bit about, you know, in the US or, uh, in Europe what people are working on. But China is both, like, running ahead on materials development-

Yes

... and also is thinking about a lot of these ideas about end-to-end labs going all the way through manufacturing, and they, they do have the expertise to do that. So what is your thought about how we stay competitive versus China, and, um, like, what is the most important thing we need to focus on there?

Yeah. This is an incredible question. This is actually what I spend a lot of my time on and when I'm in DC is just really talking about. So- China has an unfair advantage that we do not wanna replicate, but must figure out how to defend against.

So in China, they really will go out of their way, and there's incredible work by, like, NIST and a couple other groups to documenting this, to stand up manufacturing innovation hubs where they lo- they make a new material, and they will s- you know, support via capital or infrastructure the scale of said material system or, or said invention.

And they can do that because honestly, in China, whether you're public or private, one entity owns everything. And again, that's the part that we should not mimic, we should not copy. In no way am I suggesting that. What I'm saying is because they have that, we need to have a similar focus.

We need to figure out how to break the 25-year timeline that when I come in here, and I tell both of you, "Materials are long," right? You're like, "Yeah, that's literally, that's all we hear about. Everyone we talk to says the same thing.

We know the same thing." We need to change that. How do you do that? I think number one is you start to teach the scientist of the future how to run science this way. You know what the most impressive thing from our lab is?

I love everything we've talked about today. What is so impressive is we can have one PhD in metallurgy or alloys run 10 campaigns at a time. When I was in a, in a PhD, uh, we had 10 scientists focused on one campaign, one research problem at a time.

That's an order of magnitude jump in productivity from one scientist. Now imagine every scientist in the United States, every scientist in North America, in the world, doing 10 times the research output. That's fundamental. I mean, that just changes the trajectory of discovery.

But China can do that, too, right?

Yeah, they can.

Yeah.

They can. So I think the second piece is investment, and this is where I was gonna... You know, I think we're getting it in the private sector, which is great. I think you'll-- we'll continue to see the government invest in this area to try to start bridging these gaps and building up this workforce.

Um, you see the Genesis mission as one area where there's a ton of, you know, hundreds of millions of dollars of investment there. I know that groups internal to the national labs are building self-driving labs. My co-founder, Herd Seder, has one at Berkeley.

Argonne has one. I know that, uh, Ames is looking at self-driving labs. Livermore is looking at self-driving labs or may already has one. Oak Ridge has a manufacturing demonstration facility which is almost fully autonomous or semi-autonomous in nature.

So these labs are now starting to invest in the infrastructure to start to speed that gap up, to start to show you can shorten that gap. That's important. The third one is, is, I think maybe where we have to be different, public-private partnership, and this is what we talk a lot about.

This is the perfect opportunity for private enterprise to work with public research to create the greatest scientific tool a- as the, as the DOE likes to say, in the world. Why? 'Cause we have all the HP, uh, um, HPC that we need, high performance compute.

We have all the researchers that we need. We have some of the best scientists in the world at the national labs and the infrastructure from a material tooling or a science tooling perspective. And then lastly, you actually have the data.

You actually have more experimental data in all of those national labs than anywhere in the world, or, uh, we think anywhere in the world from all the science that we've run. So if you can couple all three of those things together, and you can bring in private enterprise who can help you make sense of that, help you close that system, help you tie the loop together.

Now, I think the entire national infrastructure runs in this way. The research infrastructure, whether that's corporate R&D or again, like, uh, SBIR, STTR, uh, uh, R&D, runs this way, and private enterprise runs this way. Now you've changed the fabric of all of R&D in the US.

Now you're at a place where you can compete with China, not by owning everything and doing unforced labor and, and the unethical practices that they might employ, but rather changing the mentality and the approach to how we do R&D.

That's how I think we can compete. That's the only way we can compete, I think, if we wanna move forward. If we do not do that, then they will continue to win because they will outpace us on cost, and they will outpace us on people.

And so if you try to play that game, it feels like, you know, we're gonna lose that game, it feels like. Um, but if you play the game of changing the system and, and building a better workforce and a better system to do R&D than they have, then we can beat them with, with raw output.

We often ask our guests a question that you've kind of answered now, but I wanted to ask it anyway and see if you have any more-- anything you wanna add, which is, if you could remove a bottleneck from-

Mm

... the industry, what would that be?

I don't think this is removable, so that's why it's not a good answer. But the hardest part about AI for science is that our feedback loops are long, right? I was-- thought I was-

That's fundamental. Yes.

Yeah, it's f- exactly. It's fundamental. That's why I don't know if you can remove it. Um, may- maybe there's ways to get it faster, but I'll, I'll give a perfect example. You think about math, like AI and math, right?

Mm-hmm.

You can run a lot of experiments in hours that, that will take us weeks or years to run in science. How do you get around that problem? That, that's a really hard problem to solve, and I think one answer is, is large scale automated systems.

So a perfect example, if you build a facility with thousand XRDs or SEMs, well, you can certainly build that model that I mentioned can do, you know, uh, image analysis better than anyone else in the world. So I think there are paths there to kinda leapfrog the challenge of doing fast experimentation, but it's fundamental, so it's a bad answer, uh, to the question.

If, if w- something that's not fundamental, I would have the tool providers restart their stack. Their tools are built for humans. They should build them for agents or robots. I feel like this is already happening in software.

Yeah.

I think you provide CLI, uh, um, uh, MCP.

Yep.

Like you're seeing this already happen.

That's right.

If you could do that at the infrastructure level for tooling, I think it would supercharge this industry 'cause now you don't wanna train someone on running an XRD or an SEM. You wanna train someone how to run the system that can do that, and that scales so much more effectively.

Now you don't need to get a PhD to analyze your alloys in SEM. Now you just need to focus on how I can run the system to do that exact analysis for me. That'd be transformational but would be quite expensive.

Awesome.

So.

Okay. Do you have any calls to action for AI for scientists, or let's say AI engineers?

Yeah, they can bring techniques that scientists are not aware of. I'm a perfect example of that. I am not a machine learning scientist by training at all. I'm a material scientist by training. I did my graduate work in my, in, in material science.

My first job out of my gra- of grad school was material, investing in material science. Now I run a material science comp- I'm a material scientist at heart. That's what I do. One of my other co-founders is a material scientist.

He's an academic professor in material science. But we have learned so much from the MLEs, uh, and the AI research scientists on our team 'cause they're not material scientists. And so they show up to a problem, and they don't get stuck on dendritic formation and grain boundary and, like, this, that, right?

They're just like, "Why don't you use a segmentation model for that?" It's like, "I don't know what that is." Or like, "Ah, that's not, that's not the first thought I have when I look at an SEM," but, but they do.

Yeah.

So one thing that they can do is they can actually, like, supercharge the industry with their own skill set. You know one thing I don't like about the industry? I see so many people trying to be the other thing.

Like, I, I meet ML engineers that are like, "I wanna be a material scientist," and I'm like, "Why? We need ML engineers that work at Radical." Like, you should be an ML engineer that helps us do science, and then I s- This one I see way more.

Material scientists try to be an MLE. Like, "I did a PhD in materials, did a master's in materials, and I gotta get on the AI side 'cause that's where the field's going." No, you just have to know how to use those tools to make you better at your job, right?

I don't, I don't care who can figure out the, the problem from SEM. I just wanna be able to figure it out. So there's so much cross-disciplinary work that I would, I would highly encourage ML engineers to, one, pay attention to AI for science.

Think that's already happening, actually. I don't think they need to hear that again, especially on this podcast. Uh, I was a huge supporter. But two, like, lean into your expertise. Bring a first principle perspective to the way that we do science.

We have been doing science the same way for hundreds of years or fifties of years or however old the tool is that we're running on. We're doing it for that long. You can come to that perspective and just totally change the way something works.

The field has already done that in the past, and it will continue to do that in the future, and I can tell you 100% guaranteed that we've done that at Radical AI today. And so-

So specialization.

Exactly. Bring the specialization and lean into your expertise. Don't shy away from it. Don't try to become a material scientist. Be an ML- be an MLE that works in material science.

Awesome. So what does your AI stack look like?

At a high level, kind of the AI scientist is really a, a, a multi-agentic approach. There are multiple agents that sit within what the AI scientist is. And, you know, we really have this orchestrator agent at the top that actually comes up with new hypotheses.

It has a specific way to actually test those hypotheses that that is internal, that allows us to, to kind of test if they're gonna be a good hypothesis and before we send it to the lab. But, but also going into that scientist are a bunch of other models as well, right?

We are taking in data sets, like industry standard data sets. You know, we pay for CalPhad, and we actually pull that data in so that we can use it the same way I would use it if I were a human scientist.

That's really important. We have a literature review agent that we've custom built. That benchmark is also public on our website. That actually can go and extract figures and information from scientific literature that's relevant to the hypothesis that we're making, so that's in that stack.

We have, uh, custom models built, you know, one of them called Matrix, which is on our website as well, and I would curry all the-- I would encourage all the ML engineers out there, go check this out. The, the model and the, the benchmark are available on Hugging Face.

You can find a blog post on our website about this. Matrix is incredible because Matrix is really a VLM that, you know, we have fine-tuned on Qwen that can actually go into images from the lab, experimental data, and extract scientific knowledge from it.

And so the obvious benefit that we saw in the model, which you can guess, is it gets really good at reading the experimental data, which is... That makes sense. The one that we maybe didn't see coming was by understanding that data, it gets better at being a scientist.

This is really cool, right? And so for the AI, uh, ML engineers out there, for the AI scientist, that is how you start to-- We talk a lot about this intuition, the scientific intuition you get a PhD on.

That's how you start to capture that. Now we have actually seen this, and you can go read the publication. It's on arXiv, where the public data set that we used actually is showing improvements, like five to to 16%, I believe, on general scientific reasoning.

And so we even-

So adding math to your reasoning-

Yeah. Actually, I think math's the one area we call it that doesn't work.

Okay. No, but the, the theory is the same, right?

Yes.

Where you add math and then you end up in other domains.

Yeah, so and, and we do in the paper move outside into the bio space, I believe, and, and see that same improvement. You can find this all in the preprint that's out on arXiv. So what's cool about this?

What's so cool is that as you start to build these systems that can, like, do science like a human scientist does, you start to compound the knowledge in the way we talked about earlier, where I can be looking at all of these different things as I'm making a hypothesis.

And although us human scientists would like to think we can pull in CalPhad, pull in literature, and, and extract the right information from literature, not just read a paper, but pull the right stuff out that's relevant to this hypothesis, look at all my past experiments, all of the database of our experiments are feeding into that AI scientist, so when we make a new campaign, it is looking at the past results to make that campaign.

And we have a really cool demo that we show customers where when you look at a hypothesis generated, it'll actually tell you what experiments it's pulling into, what it's using to learn from about why it made that new hypothesis.

So that's really cool. And then again, into these models like Matrix or Matrix PT, where you actually can start to pull out intuition, and then that intuition actually helps you become a, a better scientist at, at large. This is really important concept, this, this multi-agentic approach, as I think what an AI scientist really means.

I don't think it's ever gonna be one scientist. M- maybe something will happen in the future. If you're an MLE, maybe that's a good problem to tackle. Uh- ... build that, and we would love to be a customer of yours.

Um, but if not, I, I think you're gonna have these specialized models, these specialized agents that are really good at one thing, and together, collectively, they make a scientist that's better than Joseph, that's better than the scientist that we have today.

That's kinda what our stack looks like at a high level on the hypothesis generation and new material side.

And then just quick follow-up. So, um, I love that you're open sourcing a lot of your work.

Yep.

Why are you doing that? Not that I wanna discourage it in the least.

Yep. Uh, this is a really important question. Number one, there's three reasons. Number one, I talked about community in this episode. Um- We need the community to move to doing science this way. Open source work is some of the best ways to do that, as we've seen over history.

So that's number one. Number two, learning. We actually get way more feedback from open sourcing things than we could possibly work on ourself. You guys are probably aware of TorchSim, which was this package that we open sourced. We don't need to go into that today.

We can-- The, the, the feedback from the community, the ideas from the community, we've actually spun TorchSim out into its own organization, a nonprofit that can actually continue to run with the community, and, you know, we call Ignite a materials revolution, a simulation revolution, and which we're excited about.

So that's a good example. Uh, and so this, this idea that you can build better technology with the group, uh, is, is number two. Number three, we actually don't think models are remote. We actually think in five years, most models will be open source.

Yeah, there's probably a proprietary model or two, the same way there's a proprietary model or two today that I can run, whether it's Claude or, or, or, um, ChatGPT or Grok or whatever your favorite AI is. However, we think in science, models aren't the moat, experiments are.

And so we actually think the more great models we can share, like Matrix, which is out there, um, the dataset's not, right? That data's the, the model that we built and put in the preprint is built on public data.

Of course, we have our own proprietary data stacked on top of it. That's running Matrix internally. Then the model can go out there, and I, I tell people all the time, "What happens if someone else comes out with a better foundation model, better MLIP or a better diffusion model?"

I'm like, "That would be amazing." It would supercharge our scientists. We'll drop it right into our stack.

Yeah.

That's not-- We, we don't wanna sell models. We don't sell models. We think the entire community will continue to have ideas that we cannot have alone. That's why we open source because we don't think that's the edge. We think the edge is in the experimental side.

And yeah, that's specific to Radical, and obviously I'm talking about what Radical believes and the thesis there, but that is a big part of why we do it. If the whole community can push the whole field forward, we benefit from that, and they benefit from that.

That's a win-win for materials at large. That's a win for AI, for science, and it's a win because our stack now is a better model that we didn't have to produce, which is great. Uh, so whe- so whether it's us, we, we do have custom models built internally or someone else, we port that model in or use that or even pay for proprietary model, which we do on the LLM side, obviously.

We don't build custom LLMs. Of course, we use ChatGPT or, or Claude. Um, all of that just goes into making a better scientist, and that's why we open source for the community to get better ideas and, and to really understand that we think experiments or really push that we think experiments are the moat, not the model itself.

Well, yeah. Thanks for chatting.

Yes, Joseph Krause, thank you so much.

Thanks for having me, guys. Awesome to be here.

I'm glad you're, you enjoy the show so much. I hope that, that we can, you know, spread the, spread the good vibes.

You guys are, and I think we just talked about a great closing topic, which is how important this industry is. The impact that the world will feel from AI for science is enormous. People that can start at the ground root and actually push that forward, like yourself, are imperative.

So thank you for everything you do. Super happy to be here, and, uh, looking forward to coming back when the lab is fully autonomous.

Yeah.

We'll, we'll run an episode in the lab.

Yeah, no.

I was gonna say that.

Yes.

We should definitely do that. We have to do that.

When you come back to New York. Yes.

Yeah, no problem.

Yes.

Uh, pick a different time than the winter, though. We had a tough winter this year.

Okay. We'll do summer. We'll do summer.

Perfect. Thank you guys for having me.

Awesome. Thank you very much.