For two decades the bottleneck in carbon capture was not the idea. It was the material. Chemists tested sorbents one at a time, betting years of lab time on each candidate. That hunt now runs on a different engine: generative models that propose thousands of structures before a single beaker is warmed.
The shift matters because direct air capture (DAC) lives or dies on the sorbent. A better material cuts the heat needed to release captured CO2, and heat is the line item that decides whether a plant earns money or burns it. The labs that used to screen what existed are now designing what should exist.
The material was the wall.
TIMELINE: AI-designed carbon-capture catalysts
─────────────────────────────────────────────────────────────
2023 ──── 2024 ──── 2025 ──── 2026 ──── NOW
🧪 🤝 🔬 🚀 ◉
OC20/ Numat + MatterGen Zn(II)
GNoME Meta AI (Nature) catalyst
ML MOF DAC npj CO2→ + large-AI
potentials collab methanol models
descriptor
SOURCE: Microsoft Research; Numat; npj Comput Mater; Energy & Fuels (2026)
From learned force fields to generative sorbent design, 2023 to 2026.
Screening gave way to generation
The first wave was subtraction, not invention. Models such as OC20 learned to predict how a catalyst surface would behave, letting teams rank known materials without building them. Useful, but it still searched a library someone else had filled.
Generative design removed that ceiling. Microsoft Research's MatterGen, published in Nature in 2025, produces stable, novel inorganic structures aimed at a target property. The team reports 78% of its generated structures sit within 0.1 eV per atom of density functional theory (DFT) energy minima, the bar for "this could actually be made." That is the bar. Fine-tuned for CO2 capture, it becomes a sorbent foundry.
Separately, a npj Computational Materials paper in 2025 built machine-learned descriptors for CO2-to-methanol catalysts across roughly 160 metallic alloys, surfacing candidates such as ZnRh and ZnPt3 that classical screening had not flagged. Same lesson, new reaction. The model finds corners of the periodic table others skipped.
The companies already placing bets
Numat, a Chicago materials firm, paired with Meta AI's Fundamental AI Research group in 2024 to build predictive models for next-generation CO2-capture metal-organic frameworks (MOFs). The logic is pragmatic. MOFs offer the tunable pore architecture DAC wants, but the design space is effectively infinite, and experiment alone cannot cover it.
Researchers at the University of Kentucky reported zinc(II) enzyme-mimic catalysts for DAC that lift capture rates up to twofold and improve CO2 mass transfer by 40 to 60 percent in carbonate sorbents, under ambient conditions. Bench-stable and earth-abundant. They work at the dilute CO2 levels real air holds, attacking the kinetic wall that slowed carbonate capture.
The 2026 framing from the field is a new era, where large AI models, universal machine-learned interatomic potentials, and robotic self-driving labs compress discovery from years to weeks. That is the part to watch: the loop, not the paper.
Why this is an infrastructure story
Carbon capture has a poor track record of announcements outrunning commissions. As we wrote in August, Chevron built a DAC machine at the Kern site and then walked away, a reminder that engineering and offtake, not the science, often decide outcomes. Engineering beat science there. The enzymatic route we covered the same month, Carb Enzero's move from lab to pilot, shows a different path: a specific biology doing a specific job.
AI-designed catalysts are not a rival to those approaches. They are a faster way to stock the shelf both of them draw from. The investment question is whether the discovery advantage turns into deployed capacity or stays a paper lead.
First DAC sorbent validated end-to-end from an AI-generated structure, not a hand-picked one
A self-driving lab closing the predict-synthesize-test loop on carbon-capture materials
Patent filings shifting from method to specific generated compositions, the same arc already visible in polyolefin catalysts
Cost-per-tonne movement tied to a new sorbent, not a new process
Where the timeline bends
The easy milestone is another record sorbent. The hard one is a generated material that reaches a pilot and stays stable for a year. Most AI-discovered catalysts still die in transfer from simulation to real flue gas, where humidity and impurities punish ideal structures.
The second bend is ownership. Whoever holds the generated compositions and the data behind them holds the moat. Open releases such as MatterGen lower the floor for everyone; proprietary loops such as Numat's and IBM's raise one company's ceiling. That split defines the competitive map for the next three years.
What it means for the thesis
Treat AI catalyst discovery as a pick-and-shovel layer on top of the whole carbon-capture build-out, not as a capture company itself. The winners are the platforms that turn compute into verified materials and the labs that close the loop with synthesis. The losers will be the ones that publish a clever model and never ship a gram.
The technology readiness level (TRL) here is mid: credible science, early physical validation, no commercial sorbent yet born from a generative design. A mid TRL is exactly the window where thesis-stage capital gets paid, if it picks the loop over the paper.