Thesis 02 · Materials & chemistry

Inverse Design for Matter

AlphaFold showed that structure can be predicted. The next step is designing matter that doesn't exist yet, and making it.

The protein playbook is to learn structure from data, generate candidates, and test only the most promising. It's now moving to inorganic materials and other small molecules.

In 2023, Google DeepMind's GNoME predicted 2.2 million new crystal structures, about 380,000 of them stable. That's close to ten times the stable materials science had catalogued before. Then in 2025, Microsoft's MatterGen stopped searching lists and started generating to spec: ask for a band gap or a stiffness, and it designs a crystal to match. Its code and weights are open source.

Candidates are now abundant and close to free. Bringing a new material from research to first use has historically taken 10 to 20 years, and almost none of that time goes into imagining it. The constraint is testing, making and selling. Whoever closes that gap opens a market no one can size yet.

01What happens when the lab runs itself?

Self-driving labs close the gap. A model predicts, robots synthesize, instruments characterize, and the results retrain the model. Each turn of the loop makes the next one smarter.

How a new material gets found
PredictSynthesizeCharacterizeRetrainWeeksper turn

Each turn takes weeks, and every result makes the next prediction better.

The early proof is real, and still contested. Berkeley Lab's A-Lab made 41 of 58 target compounds in 17 days with almost no human hands. Outside chemists then argued that few of them were truly new. Both points matter: the loop runs, and the hard part is proving what came out of it.

The data matters more than the model. Public databases are full of idealized, computed structures. They don't record the oven temperature, the impure precursor or the run that failed. A lab that logs every turn, failures included, builds a dataset no one can scrape. As IQ Capital puts it, self-driving labs are infrastructure, not just intelligence.

Capital is following that logic. Periodic Labs raised a $300M seed to build autonomous labs, and Lila Sciences has raised about $550M for “AI science factories.”

The field is shifting from finding materials to specifying them: describe the properties you need, then design backward.

02What does this look like in production?

MatScience AI is a Harvard–MIT team that combines AI, researchers in the loop, lab validation and manufacturing to develop product-ready materials. Its flagship product, already in production, is a PFAS-free liquid coolant for high-density AI data centers.

According to the company, it targets savings of roughly $2–4M per gigawatt per year through better heat transfer, lower pumping power, less maintenance and a longer service life. The same stack is now working on magnesium alloys for semiconductor and aerospace uses.

The point isn't the coolant. It's the wedge. AI racks are hitting a thermal wall now, and regulators are pushing PFAS out of the fluids many cooling systems rely on. That's two deadlines on one buyer. A model-designed material aimed at both skips the years of “interesting, let's talk next year.”

03Where does it break?

In software, a working prototype is most of the journey. In materials, a gram in a vial is the start. Most companies here will stall between the lab and the customer, in three places:

  • Scale-up: heat, mixing and impurities behave differently at a kilogram than at a milligram, and again at a ton.
  • Cost: a model can propose a crystal that needs exotic precursors or extreme conditions that no plant can run at a profit.
  • Qualification: buyers retest everything before it touches a production line. In automotive alone that's often 12 to 18 months, and it's paid for out of the startup's runway.

Even the best-funded platforms are still crossing this gap. Coverage of CuspAI's $450M round described AI materials discovery as entering its validation phase.

The winners will own the commercial path: who buys it, at what price, and how long qualification takes.

04Who's building it?

The field is splitting into three models:

  • Material makers: own the loop and sell the material. Highest control, most capital. MatScience AI fits here, and so does Materials Nexus, which developed a rare-earth-free magnet in three months.
  • Science factories: build robotic labs at scale, generate proprietary data, and license or co-develop with industrial majors. Periodic Labs, Lila Sciences and CuspAI.
  • R&D software: help incumbents get more from the data they already have. Fastest to adopt, lowest ceiling. Citrine says its platform cut one customer's formulation time from six months to one.

My bet is on the first model, and on factories that pick a product early.

05How big is the market?

Like the biomarker thesis, this one has no TAM yet. You can't size a market for materials that haven't been invented: rare-earth-free magnets, better battery electrolytes, coolants for chips that don't exist yet.

What you can size is the wedge. Data-center cooling, packaging materials and specialty catalysts each have a buyer with a budget and a deadline. Start there, and treat the platform as upside.

What I look for

  • The loop: do they own design, validation and manufacturing, or depend on outside labs for every result?
  • First product: is it aimed at a buyer whose roadmap is blocked today, like data-center cooling, batteries or semiconductors?
  • Data: does every loop turn log conditions and failures that can't be scraped from papers or public databases?
  • Path to scale: is there a process engineer on the team or a manufacturing partner signed, and how long does the buyer take to qualify it?

What would make me wrong?

  • If platforms that license software to chemical majors win instead of companies that make materials themselves, the slow path to scale matters less than I think.
  • If lab robotics gets cheap and standard enough that any mid-sized chemical company can run its own self-driving lab within a year, the startups' data advantage disappears.

Qualifications

ILoop

How fast does discovery actually turn?

  1. 1.Which steps of predict, synthesize, characterize and retrain do you own?

    Owning the full loop is the moat; what is the path to getting there?

  2. 2.How many loop turns do you run per month today, and what limits it?

    Every turn teaches the model something. The team that runs the most turns learns the fastest, and that speed is what this thesis is betting on.

  3. 3.What in your experimental data can't be reproduced from public databases?

    Models commoditize. MatterGen is already open source. Proprietary measurements and failed runs don't commoditize.

IIProduct

Is there a buyer who needs this now?

  1. 4.What's the first material, what problem does it solve, and what does it set up next?

    The first product should fix something a buyer already measures and pays for, like heat, cost or cycle life. The best ones also leave behind the data, customers or production line that make the second material easier.

  2. 5.Who's the buyer with urgent demand, and what does qualification involve?

    Qualification can run a year or more, and it's the hidden runway cost in materials.

IIIScale

Can it be made, and at what margin?

  1. 6.Who manufactures at volume, and what do precursors cost at ton scale?

    Lab-scale wins often die at kilogram and ton scale.

  2. 7.Do you sell material, license designs, or co-develop with a partner? What are the margins on each?

    This decides whether the slow path to scale is your problem or your customer's.

IVDefensibility

What's still yours in five years?

  1. 8.If open materials models get as good as yours, what remains defensible?

    Assume the model is a commodity and see what's left: data, process know-how, or a buyer's qualification.

  2. 9.Which incumbents could buy or build this, and why haven't they?

    Chemical majors move slowly until they don't.

If you're building at this edge, I want to hear from you.

Tell me what you're designing and who's waiting for it. I read every submission myself and connect the strongest founders with the investors and operators who should know them.

Next: Thesis 01 · Biomarker Expansion →
Raising now?

140 characters max

Sources

  1. 1.Merchant et al., “Scaling deep learning for materials discovery” (GNoME), Nature, 2023. nature.com
  2. 2.Zeni et al., “A generative model for inorganic materials design” (MatterGen), Nature, 2025. nature.com
  3. 3.Materials Genome Initiative white paper, OSTP, 2011. mgi.gov
  4. 4.Szymanski et al., “An autonomous laboratory for the accelerated synthesis of inorganic materials” (A-Lab), Nature, 2023. nature.com
  5. 5.Chemistry World, “New analysis raises doubts over autonomous lab’s materials ‘discoveries’,” 2024. chemistryworld.com
  6. 6.IQ Capital, “Self-driving labs = infrastructure, not just intelligence,” 2026. iqcapital.vc
  7. 7.TechCrunch, Periodic Labs $300M seed, 2025. techcrunch.com
  8. 8.Fierce Biotech, Lila Sciences Series A, 2025. fiercebiotech.com
  9. 9.MatScience AI, MIT Orbit listing. orbit.mit.edu
  10. 10.Schneider Electric, “PFAS phase-out and liquid cooling: What US data center operators must do now,” 2026. blog.se.com
  11. 11.I-Connect007, Indium on automotive material qualification, 2018. iconnect007.com
  12. 12.eWeek, “CuspAI Raises $450M as AI Materials Discovery Enters Its Validation Phase,” 2026. eweek.com
  13. 13.Henry Royce Institute, Materials Nexus case study. sheffield.ac.uk
  14. 14.Citrine Informatics, Dorfner case study. citrine.io