A space too big to search

We work on metal-organic frameworks or MOFs, which won the Nobel Prize in Chemistry in 2025. 

They are highly tunable materials made of metal nodes connected by organic linkers. They are mostly known for their large surface area: one gram can have the internal surface area of a football field. This is why MOFs have no equivalent in gas and liquid separation, from capturing CO2 to harvesting water from air, storing hydrogen and sensing chemicals.

Machine learning and generative models have changed the game and completely changed the pace of discovery for these materials. In practice they run into two problems. The chemical space of possible MOFs is enormous, and there might be more MOFs than stars in the Milky Way or galaxies in the observable universe. And most of what a model generates cannot actually be made. The second problem is the one that matters most. A material is only useful if it can be synthesized, otherwise it just stays on a computer.

We wanted a system of agents and diffusion models, trained on the whole synthesizable MOF space, that took both problems seriously at the same time.

Why building a system of agents?

Our first instinct was to train a single large model and we found that most of the materials generated could not be synthesized in the lab. We gave that up quickly.

"Designing a new material" is not a single task, it is a set of interconnected tasks that each need a different kind of intelligence. Proposing a chemically meaningful pairing of metal and molecule is a knowledge problem. Placing every atom in three dimensions is a geometry problem. Checking that the crystal does not break the laws of quantum mechanics is a physics problem. And judging whether a chemist at a fume hood could actually make the thing is a synthesis problem, by a large margin the hardest of these tasks.

So instead of one model we built a system of specialized agents each responsible for one question, or task, screening candidates at each step. We called it MOFGen.

A large language model proposes compositions based on the basic chemistry learned from the MOF literature. With Google DeepMind we trained a diffusion model, the same family of model that generates images from noise, on the whole periodic table and already-synthesized MOF space to build the crystal structure atom by atom, scaling up to 256 atoms per unit cell. This is one of the central pieces, since conventional generative models for crystals stop at a few dozen atoms, closer to the size of a molecule than of a crystal. Quantum mechanical agents relax and filter what comes out, discarding anything unstable. Finally, synthesis-feasibility agents ask the question: can this be made in a wet lab?

Running the full pipeline, we generated hundreds of thousands of MOF structures, making it one of the largest computational MOF databases to date.

The uncomfortable question

If there is one thing I would want a reader to take away, it is that a generated structure is not a material. It is a hypothesis. Generating a new material is easy: some of them look beautiful and have incredible predicted performance, but nobody can make them.

That is why the synthesis and quantum mechanical agents matter as much as the generative ones. Much of our internal debate was about how strict those filters should be. Filter too hard and you discard everything interesting. Filter too loosely and you hand your experimental chemist a list of impossible tasks.

Four is the number that counts

Hundreds of thousands of structures generated. Eight materials that crystallized but only four fully characterized with their crystal structure.

We narrowed the candidates through high-throughput experiments. Tens of candidates were tried, eight crystallized, and only four "AI-dreamt" MOFs were fully characterized. When our article was first released, that was the largest set of AI-designed materials anyone had carried from a generative model to a real material.

What surprised us

The diffusion model has a bias we did not expect. It tends to produce flexible frameworks with "breathing" behaviour as guest molecules move in and out. For some applications that is critical, but for others it is precisely what you do not want. A generative model learns whatever is easiest to learn from its training data, not what you had in mind.

Another honest limitation: the 256-atom limit constrains how large a framework we can generate, and MOFGen is not autonomous. A human is still in the loop for the synthesis, deciding what is worth trying given finite laboratory time and money.

Where we go next

MOFGen is modular by design, so each agent can be improved or replaced without rebuilding the system. What interests us, and what we believe is the future, is generating frameworks targeted at specific properties, as done for inorganic crystal by Microsoft with MatterGen, adding feedback so that experimental outcomes update the generators and connecting the output directly to self-driving laboratories.

The long-term goal is a materials discovery laboratory that closes the loop on its own. We are not there, but we are not far, and much closer than readers think. Four fully AI-dreamt frameworks now sitting in vials, more than any previous study has carried from a generative model to a real crystal is an argument that this path exists and that the next generation of materials we need for carbon capture, water harvesting and hydrogen storage might be found far faster than we think.