AI Goes Deep

AI Goes Deep

By FedTechBusinessTechnology
Download on the App Store

All AI Goes Deep episodes

  • Same Recipe, Different Cake: Why Lab Data Isn't Ground Truth

    Same recipe, two kitchens, two different cakes — and nobody wrote down the oven. That's the problem with most lab data, and why AI can't learn from it. Dr. Jehad Abed, materials scientist at X, The Moonshot Factory and former experimental lead on Meta's Open Catalyst Project, explains what AI needs from the lab instead.

    We cover why stitching other people's datasets together doesn't give you ground truth, what a catalyst does (turning waste CO2 into jet fuel, for one), how AI made chemistry simulation cheap enough to model real-world defects, and why self-driving labs need shared standards. Then the part founders should hear: discovery isn't where materials companies die. Scale-up is.

    Jehad, on what the field gets wrong: "We tend to celebrate material discovery more than we celebrate process discovery."

    **Guest:** Dr. Jehad Abed — Materials Scientist at X, The Moonshot Factory (Google). More than a decade at the lab bench; formerly Meta FAIR, where he led experiments for the Open Catalyst Project.

    **Mentioned:**

    — Open Catalyst Project and the OCx24 experimental dataset
    — ARPA-E's CATALCHEM-E program for self-driving catalyst labs
    — Alán Aspuru-Guzik and the self-driving lab community
    — Titanium, and why a better way to make it mattered more than discovering it
    — Episodes 1 (James Warren, NIST) and 2 (Andrey Ivankin, Mattiq)

    A podcast from FedTech.

    37 min
  • Materials AI Has a Reality Problem

    Materials AI has a reality problem: the models are trained mostly on simulations, and most of what they predict never gets made. Dr. Andrey Ivankin, co-founder and CTO of Mattiq, is building the machine that produces real experimental data at scale.

    Mattiq prints hundreds of thousands of unique nanomaterials onto a single chip, each positionally encoded, then screens them. It's the genomic chip, brought to inorganic materials. We cover what that does to the discovery loop, where characterization gets slow, why the previous generation of materials informatics companies couldn't generalize beyond one vertical, and what Andrey would tell a technical founder who's thinking about leaving the lab.

    Andrey's take on the state of the field: "The training data is mostly computational, most predictions never validated. We need to change that."

    Guest: Dr. Andrey Ivankin — co-founder and CTO of Mattiq, co-founder of TERA-print, chemical physicist and engineer. Mattiq has raised close to $25M in venture capital plus government funding and is currently raising a Series A.

    Mentioned:

    — Polymer pen lithography, and Chad Mirkin's lab

    — Scanning droplet electrochemical cell; AFM; STEM and EELS

    — The Inorganic Crystal Structure Database — roughly 250,000 materials

    — Google DeepMind's GNoME; Lila, Periodic Labs, Radical AI

    — The Genesis program

    A podcast from FedTech.

    35 min
  • Self-Driving Labs and 15 Years of the Materials Genome Initiative

    Dr. James Warren has directed the Materials Genome Initiative since 2012. Fifteen years after the US government set out to halve materials development time, he explains what landed, what didn't, and where the door is for founders.

    We cover why materials science is really just cooking, when AI is worth reaching for, what a self-driving lab does that a supercomputer can't, and why a 7:1 return on investment still isn't enough to move a manufacturer that's already profitable.

    Warren's argument, in one line: adoption isn't a question of if. "Whether we have to go through a full Gartner hype cycle on this, I don't know. What I do know is that this is inevitable."

    Guest: Dr. James A. Warren — NIST physicist since 1992, Director of the Materials Genome Initiative, co-founder and director of the NIST Center for Theoretical and Computational Materials Science.

    Mentioned:

    — Warren's paper, "Overcoming Roadblocks for Implementing AI/ML Methods for Materials Advancement" (Annual Review of Materials Research, 2026)

    — NSF's programmable cloud labs: $400M, 20 teams, four years

    — DOE's Genesis Mission

    A podcast from FedTech.

    30 min