.avif)
Google and Microsoft have both shipped models that predict new materials. Years on, most of those predictions have never actually been made in a lab. Dr. Andrey Ivankin, co-founder and CTO of Mattiq, calls that a reality problem — and he is building the hardware to fix it.
Watch the full episode right here:
Listen instead: Spotify · Apple Podcasts · YouTube
The short version:
- Materials AI is trained mostly on computation, and most of what it predicts is never made. That gap is the whole problem.
- Synthesis used to be the slow step. Put a hundred thousand materials on one chip and it stops being the slow step — characterization becomes the constraint instead.
- The previous generation of materials discovery companies failed by narrowing. Mattiq's bet is that going broad is what makes the model generalize, and the chip is what makes going broad affordable.
Episode one was the view from inside the federal system — Dr. James Warren at NIST, fifteen years into the Materials Genome Initiative. For episode two I wanted the view from the bench: someone actually generating the data everyone says the field is missing.
My guest is Dr. Andrey Ivankin, co-founder and CTO of Mattiq. He's a chemical physicist and engineer by training — "a rare and proud Siberian," as he puts it. In 2015 he co-founded TERA-print with Professor Chad Mirkin, building nanofabrication tools now used by researchers around the globe. Mattiq came out of the realization that those tools, combined with new chemistry from Mirkin's lab, could change how materials get discovered in the first place. The company has raised close to $25M in venture capital plus a few million in government funding, led by Material Impact, and is currently raising a Series A.
This one is personal for me. I spent six years of my PhD using polymer pen lithography, the printing technology underneath all of this. I can testify the technology works.
Editor's note: Quotes are condensed for length and clarity. Questions and answers have been grouped by topic rather than in the order they occurred. The full transcript is here. [TRANSCRIPT LINK TBD]
Why can't we discover materials fast enough?
Because materials science never built the machinery to run experiments at scale, and biology did.
You want better batteries, you want more energy efficient AI chips, you want to explore deep ocean, or you want to go to space — all of those technologies require new materials, and we just can't discover them fast enough.
The reason for that is that unlike in biology, materials science never developed experimentation at massive scale. We still live in this pre-genomics-chips revolution in materials science. And without that scale, you can't explore design spaces faster. You don't have experimental data at scale that you can use to train AI that will be consequential, and you can't rapidly improve that AI to accelerate the discovery. So we are stuck in that slow, expensive discovery cycle.
The comparison to biology isn't decorative. Andrey spent time in biotech and DNA-related research, and he's describing a transition he watched happen once already:
First there was serial experimentation, and then people started to do a lot of automation in biology — 96-well plates, 384-well plates, adding robots to run experiments faster. But the unit of experiment was still one well at a time. It's still slow, and to scale your experimentation in that case you need more robots. The throughput is not easy to scale.
When genomic chips arrived, now you have a chip with, instead of 96 experiments, 10,000, 100,000, millions of experiments on a single chip. And how do you scale your throughput? You just reduce the density. So you can run much more, it becomes cheaper, faster. That's what we are trying to bring to materials science for the first time.
What is a megalibrary?
A single chip carrying hundreds of thousands of different materials, each one in a known location.
I like to think about it as genomic chips for inorganic materials. We brought together massively parallel printing — a technology called polymer pen lithography — and combined it with innovation in chemistry that allows us, on one chip, to synthesize hundreds of thousands of unique nanomaterials, all positionally encoded. Then we screen them for any properties of interest, from structure to function. Whether it's a good magnet, whether it's a good conductor, some other type of battery material — at a pace you couldn't do before.
For anyone who hasn't held one of these: it's roughly a square centimetre, patterned with hundreds of thousands of tiny pyramid-shaped tips. You engineer the chemistry and the composition, print once, and you have a vast relationship database — these conditions and this composition produce this result — laid out in a grid you can go back and read.
I asked him how fast, in practice. His answer separates the two halves of the problem, and that separation turns out to be the most important thing in the episode:
We should separate synthesizing materials and testing them. In terms of synthesis, we can print chips in an hour, maybe even less. So in one day with one instrument you can print ten of those chips.
Characterization really depends what properties you go after. Some techniques — optical, or some electrochemical — can run very fast. In a matter of a few hours you can characterize a hundred thousand new materials. Structural characterization, some additional functional characterization, those can be slower, and you need to be wise about how you characterize. But nevertheless you can run at a much faster pace — orders of magnitude — than with any alternative approach.
How is this different from a self-driving lab?
It changes the shape of the loop, not just its speed.
Self-driving labs — the subject of most of episode one — close the loop between computation and experiment: an AI proposes the next experiment, machines run it, the result feeds back. Andrey is careful that Mattiq is inside that category rather than opposed to it. But the loop itself changes when synthesis stops being a step you wait on:
The traditional loop is: generate hypothesis, synthesize one material, test it, learn something from it, repeat. That loop can take, in a traditional automated workflow, hours — in some cases days per material. And which step is the longest? In most cases it's actually synthesis.
With a megalibrary, you synthesize a hundred thousand materials at the same time, positionally encoded, and your loop becomes: generate hypothesis, go measure, learn, refine hypothesis, repeat. So you eliminate the longest step from your loop almost completely — I mean, you still need to synthesize new megalibraries occasionally. But that gives you the ability to test not one hypothesis at a time, but hundreds or thousands of hypotheses at the same time.
There's a real trade being made here, and it's worth naming plainly. Delete synthesis from the loop and you don't arrive at a solved problem — you arrive at a different bottleneck.
The bottleneck moves to characterization
Making the material was always half the job. The other half is proving the thing you made is the thing you meant to make. Mattiq started with the techniques that read fastest:
Obviously you try to start with something easy and simple. In that regard we started with optical characterization and electrochemical characterization — a technique known as scanning droplet electrochemical cell. It's a small droplet where you run your chemical reaction, and you can raster that droplet across the chip and measure the activity of a given catalyst at every single point. We started there because those techniques are easier and faster to produce data. But that's not where we want to stop.
Where they want to go is harder and, he argues, more general. Atomic force microscopy is the near-term expansion, in partnership with instrument makers:
We work with Bruker and Oxford on expanding and developing the characterization and application of that technique to megalibraries. You will never have the throughput of optical microscopy, but you can run fast enough that it becomes very meaningful.
Atomic force microscopy I like to call the Swiss knife of characterization, because it can be operated in so many different modes. You can look at whether your materials are good magnets, whether they're good conductors, at mechanical or thermal properties, piezoelectric properties. All of that information becomes available.
And structure, which is the part that makes a material genuinely characterized rather than merely interesting:
You only have a material fully characterized when you have composition linked to synthesis, linked to structure, and linked to the properties that material produces.
Then the number that stopped me:
If you think about what's available today — humanity to date has characterized and documented in the Inorganic Crystal Structure Database around two hundred fifty thousand different inorganic materials. When you combine megalibrary and STEM in an automated fashion, you can surpass that in a matter of a week with one instrument.
Worth adding a footnote of my own here: the ICSD sits behind a paywall. Meta and Google have both open-sourced materials databases and described them as ground truth, but those are computational results. There have been cases where structures announced as discoveries were already documented in the literature — plausibly because the paywalled record wasn't available to the teams involved. The access problem and the reality problem are tangled together.
So what's the business model?
Not selling data. Selling the judgment the data produces.
Data by itself is not very useful. Data can provide insights, and the more data you have, the more generalizable your insights become across materials and across properties.
That word — generalizable — is where his read on the industry's history sits:
Our thesis is that most materials discovery companies failed because they couldn't generalize. What happened is they had to focus on a very specific vertical and a very specific application, and that doesn't scale very well. Our premise is that by going at scale and broad across applications and properties, eventually we have a model that can get us answers across applications fast.
And the beauty here is that we don't only have that model, we have the engine, the real-world engine behind it, that allows us to quickly validate those predictions and refine the model. So we arrive at the answer in very few loops.
In the near term that means a few verticals — catalysis and energy more broadly, magnetics, semiconductor materials — and partnerships rather than product sales:
We partner with companies who have materials problems, because our data has two benefits. It helps you find new materials faster, and it also provides guidance on which materials are easier to scale and bring to market.
We engage, share our capabilities, make sure we operate in the same domain in terms of materials and properties of interest. Then we define the problem statement — what material you're trying to identify, what you're trying to solve for. We define the milestones, the budget, and the capacity of the factory required to explore the universe you're interested in, and see if we can discover completely new materials and generate IP that's not out there. Once we generate validated materials on chip, we pass along the formulations and you validate.
We don't think we replace your R&D necessarily. But if you can do hundreds of thousands of experiments a year, we just give you the best chance that you'll succeed with those experiments and focus on the best materials.
The second benefit is the underrated one. A model that tells you which candidate is easiest to manufacture is doing something a materials model usually doesn't: it's answering a question the person with the budget actually has.
Why did the newcomers raise hundreds of millions?
I've tracked materials informatics for a long time, and suddenly there are newcomers raising hundreds of millions — at least an order of magnitude more than companies like Mattiq. I asked him what those companies are doing differently. His answer is the argument at the centre of the episode:
One interesting shift in the past couple of years: if you look at the GNoME model that Google DeepMind came up with in 2023, and then Microsoft came up with some models — everyone was excited, but the reality is we are many years later and most of those predictions have never been validated. If you look at DeepMind, they're building their own wet lab to run materials experiments. The field is realizing that we need experimental ground truth to train models and actually make something useful out of them.
Not to say that the models are bad. But I like to say that materials AI today has a reality problem. The training data is mostly computational, most predictions never validated. We need to change that.
The field understands the problem, he says. The difficulty is what fixing it costs:
If you try to do it by automating experimentation the way we've been doing it to date, you require a lot of CapEx and a lot of development. And not only development — because the whole field is not ready. Instrumentation vendors, the instruments themselves, are not ready for running in the loop like a self-driving lab. So you need to put a lot of effort into optimizing and developing APIs and so on. But also, to get to a meaningful scale, you just need a lot of CapEx if you're doing it the traditional way.
Which is a fair description of why those rounds are the size they are. Building physical reality into the loop is a capital-intensive act, and someone has to do it.
He doesn't frame the well-funded labs as competitors:
Going back to the analogy with biology — we don't see megalibrary as an alternative approach to what Lila or Periodic Labs or Radical AI are building. We operate at different scales. We work at small, nano scale, and the advantage we get from that is we can run much faster and more cost-efficiently. At the same time, what they do is a little bit closer to the end application and the format of materials, how it's going to be synthesized. So we think about it as a funnel.
A funnel is a more useful mental model than a race. The upstream stage searches an enormous space cheaply; the downstream stage takes the survivors toward something manufacturable. Both are necessary, and they fail differently.
On leaving the lab
The last stretch of the conversation is for anyone at a national lab or a university thinking about starting a company — which, given who listens to this show, is a lot of you.
When you become a founder, you think you'll keep developing — especially coming from a technical background. You think you'll spend all your time developing the most exciting technology yourself, writing code or working on hardware in the lab. The truth is there are a lot of things you have to do and worry about. Fundraising, hiring, interacting with VCs and your supporters. All of those are very important and necessary, and I don't necessarily say that I dislike it. What I'm saying is that the scope of your responsibilities expands very quickly from you being a grad student or postdoc.
In a way there's some similarity with a professorship, because again you have to worry about bringing in students, getting funding, making your department happy. So there are more glamorous and less glamorous things you'll have to do. But at the end of the day it's very fast-paced, it's very rewarding intellectually, and if you're an idea guy or girl who always comes up with many ideas — that's the best thing to do.
His practical advice comes from having had an unusually soft landing himself:
I was lucky when I started, because my co-founder Chad Mirkin was a serial entrepreneur by then. He had a lot of knowledge and we had a lot of resources to get us started. But I see many people face a lot of challenges when they're all first-time entrepreneurs and don't have the network and support they need to start a company. That can be slow, unproductive and painful. In that sense, finding experienced advisors or co-founders early on is very helpful. You can save a lot of pain for yourself.
My own version of that, for what it's worth: ask yourself honestly whether hiring people and worrying about fundraising are things you'd be willing — or even happy — to do, rather than things you'd tolerate. If the answer is no on enough of them, that's not a reason to stay in the lab. It's a reason to go find a co-founder.
Three things I took away
- The reality problem is the whole thesis. Computational training data with unvalidated predictions is a self-reinforcing loop, and everyone serious is now spending real money to break it — DeepMind with a wet lab, Mattiq with a chip, the newcomers with nine-figure rounds.
- Deleting a bottleneck relocates it. Take synthesis out of the loop and characterization becomes the constraint. Mattiq's roadmap is really a characterization roadmap, which is a more honest way to read the company than the AI framing.
- Generalization is a strategy, not a property. The claim that earlier materials discovery companies failed by narrowing into one vertical is a specific, falsifiable bet. Going broad is expensive, and only affordable if your per-experiment cost is low enough — which is what the chip is for.
Mattiq is at mattiq.com. They're raising a Series A and open to conversations with partners who have materials problems — Andrey's LinkedIn is here. Full episode transcript here. [TRANSCRIPT LINK TBD]
Episode one, with NIST's Dr. James Warren on fifteen years of the Materials Genome Initiative, is here.
If this was useful, follow along wherever you listen — and send it to the person you think will argue with us.
AI Goes Deep is a podcast from FedTech.
