Climate · Artificial Intelligence

The Molecular Sponge

To pull carbon from thin air, we need a material that doesn't yet exist. Machine learning is now searching millions of candidates to find it.

July 1, 2026 Lisa Pedrosa 8 min read Climate
CO₂ @ 400 ppm

The hardest thing about pulling carbon dioxide out of the sky is not the sky. It is the arithmetic of scarcity. In the open air, CO₂ makes up just about 0.04 percent of what you breathe — four molecules in every ten thousand. To capture it, you need a material that can reach into that thin soup, seize the rare carbon molecule, ignore the overwhelming water and nitrogen around it, and then let go again on command so it can be used over and over. That material, in the form we truly need, does not yet exist. Increasingly, the job of finding it belongs to a machine.

Direct air capture — DAC, in the shorthand of climate engineers — is the technology of last resort and long horizons. The world is not going to hit its temperature targets on emissions cuts alone; every serious climate model now assumes that sometime this century we will have to physically remove billions of tonnes of carbon already in the atmosphere. The problem is that today's DAC plants are absurdly expensive, in part because the sorbents they rely on are energy-hungry and imperfect. The whole enterprise hinges on a chemistry problem: find a better sponge.

~400 ppm
CO₂ concentration in ambient air
Millions
Possible MOF structures to screen
Gigatonnes
Annual removal climate models assume
Low heat
Energy to regenerate the ideal sorbent

A crystal with a hole in it

The leading candidate for that sponge is a class of materials with an almost science-fictional name: metal-organic frameworks, or MOFs. Picture a crystalline scaffold built from metal nodes joined by organic struts, riddled with regular, tunable pores — a molecular sponge whose holes can be sized and lined to grab one kind of molecule and shrug off the rest. A single gram of some MOFs has an internal surface area larger than a football field. Change the metal, change the linker, and you change what the framework catches and how tightly it holds on.

That tunability is the promise and the curse. The number of MOFs you could plausibly build is effectively unbounded — combinatorially vast, running into the millions of hypothetical structures. No laboratory could ever synthesize and test them all; a single careful experiment on one candidate can take days. This is exactly the kind of needle-in-a-cosmic-haystack search where machine learning has, in other domains, proven transformative — and where it is now being pointed at the climate.

A gram of the right crystal has the surface area of a football field. The problem is choosing which crystal, out of millions that could exist.
— On the MOF search space

Teaching the model to feel a molecule

The breakthrough of the past two years has been the arrival of machine-learning force fields — AI models that predict, quickly and accurately, how a CO₂ molecule and a water molecule will interact with the interior of a given framework. Traditionally, calculating those interactions required quantum-chemistry simulations so costly that screening a large library was out of reach. A well-trained force field approximates the same physics thousands of times faster, letting researchers computationally test enormous numbers of candidate materials before anyone touches a beaker.

Crucially, the field has learned to share. Georgia Tech and Meta released an open dataset — the Open DAC project — built specifically to train AI models on the CO₂-and-water behavior of MOFs, seeding an entire research community with a common foundation. Newer efforts fine-tune general chemistry foundation models specifically for the water-and-carbon problem, and generative systems like the recently described AtomMOF have begun predicting the three-dimensional structure of how a captured molecule sits inside its host — a detail that determines whether a framework will actually work.

The bottleneck was never imagination. Chemists could dream up frameworks endlessly. The bottleneck was time — the days it took to evaluate each one. AI collapses that evaluation from days to seconds, and suddenly the whole haystack is searchable.

Why water is the villain

If you want to understand why this problem resisted brute force for so long, look at water. Ambient air is humid, and water molecules are chemically clingy — they compete with CO₂ for the same binding sites inside a framework, and they can degrade the material outright. A sorbent that captures carbon beautifully in a dry lab can fail completely in the moist reality of the outdoors. Any honest AI screen has to model CO₂ and H₂O together, which is precisely why the newest datasets and force fields make that pairing their central task. It is a reminder that the machine is not conjuring magic; it is doing careful, unglamorous physics at a scale humans cannot match.

THE AI SCREENING FUNNEL Millions of hypothetical MOFs ML force-field screen (seconds each) Top candidates Lab test
AI collapses an unsearchable space of candidate frameworks into a shortlist a lab can actually synthesize and verify.

The honest ledger

It would be a disservice to oversell this. A promising material on a screen is not a deployed climate solution. The gap between a MOF that scores well in simulation and one that can be manufactured cheaply, packed into an industrial contactor, and cycled millions of times without breaking down is enormous — and it is a gap that AI, for now, does not close. Nor does better DAC absolve anyone of the far cheaper, far more urgent work of not emitting the carbon in the first place. Carbon removal is a supplement to emissions cuts, never a substitute, and the fossil-fuel industry has a long history of invoking future capture technology to justify present-day pollution.

Machine learning can find the molecule. It cannot excuse us from the harder work of not making the mess in the first place.
— On the limits of technological rescue

But held in proper perspective, the shift is real and it is hopeful. For the first time, the search for a next-generation carbon sponge is not rate-limited by human hands and quantum-chemistry supercomputer time. It is being run at the speed of inference, across a design space no chemist could ever have explored by hand, with the whole community pooling data instead of hoarding it. The material that finally makes direct air capture affordable may not be invented in the traditional sense at all. It may be found — surfaced from a list of millions by a model that learned, molecule by molecule, what it feels like to hold carbon in the air.

If it works, the quiet irony will be worth savoring: the same artificial intelligence whose data centers are straining the power grid may also help design the technology that cleans up after us. The race between those two facts — AI as carbon problem and AI as carbon solution — is one of the defining contests of the decade, and it is only just beginning.

Sources

  1. Georgia Tech Research — "Georgia Tech and Meta Create Massive Open Dataset to Advance AI Solutions for Carbon Capture." research.gatech.edu
  2. Chemistry World — "Machine learning helps scientists identify promising metal-organic frameworks for capturing carbon dioxide from ambient air." chemistryworld.com
  3. Matter (Cell Press) — "Accelerating CO₂ direct air capture screening for metal-organic frameworks with a transferable machine learning force field." cell.com
  4. arXiv — "The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture." arxiv.org
  5. arXiv — "The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture." arxiv.org
  6. arXiv — "AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction," Feb 2026. arxiv.org
  7. Journal of the American Chemical Society — "Architecting Metal–Organic Frameworks at Molecular Level toward Direct Air Capture." pubs.acs.org
  8. ScienceDirect — "Accelerating CO₂ direct air capture screening for MOFs with a transferable ML force field." sciencedirect.com
  9. PMC / NIH — "The Open DAC 2023 Dataset and Challenges for Sorbent Discovery in Direct Air Capture." ncbi.nlm.nih.gov
  10. ChemCopilot — "Self-Driving Labs: The Rise of Autonomous Chemical Discovery in 2026." chemcopilot.com
Ko-fi Buy me a coffee
Scroll to Top