The computer that lies. It designed a real crystal

The machine that hallucinates, and the machine that discovers

In 2023, American lawyer Steven Schwartz filed a court document full of rulings that had never been made. ChatGPT wrote them for him. Judge Kevin Castel fined him, and the world learned the word "hallucination" in a new sense: artificial intelligence makes things up — and does so with a straight face.

Two years later, another AI model proposed a crystal that existed in no database on Earth — among the 850,000 structures accumulated over decades of chemists' work. Chemists in a laboratory in China synthesized it. The crystal exists. You can hold it in your hand.

These two facts should not be true at the same time. A machine that hallucinates, and a machine that discovers. And yet.

This is the story of how AI's greatest flaw became its greatest strength.

Why finding a new material took entire generations

For a hundred years, materials chemistry relied on trial and error. Thomas Edison tested thousands of lamp filaments before landing on carbonized bamboo. On average, twenty years pass between discovering a new material in the lab and seeing it on a store shelf. The reason is simple: the space of possible crystals is astronomically vast, and human hands number only two.

The largest databases — Materials Project and Alexandria — together hold just over a million described structures. It sounds impressive. Yet it is less than a drop in the ocean: the number of potentially stable inorganic compounds is estimated at dozens of digits, on the order of 10 to the power of 60. For comparison, there are about 10 to the power of 80 atoms in the observable Universe. We know barely a million.

The classical approach to materials design is screening: you search a million known structures and pick the best one. Except screening cannot move beyond what we already know. It will not find a material that no one has ever written into a database. And those are exactly the materials we need — better batteries, cheaper catalysts, magnets without rare-earth elements.

From noise to crystal

In January 2025, a team from Microsoft Research published a model called MatterGen in Nature. Its idea is astonishingly simple once you understand one thing: the diffusion model.

Imagine a blurred ink stain on a page. A diffusion model is a machine that learned to reverse the blurring — step by step it removes noise until a sharp image emerges from the chaos. That is how image generators work. MatterGen does exactly the same thing, except instead of pixels it de-noises a crystal: it successively guesses which atoms, in which positions, and in which unit cell should be found. From pure noise. From zero.

The lattice is the key. A crystal is not a random lump — it is a unit repeating to infinity, like a pattern on wallpaper. The model must respect that symmetry, otherwise it will generate a structure that does not exist in reality. Microsoft trained MatterGen on 607,683 stable structures from the Materials Project and Alexandria databases, then let it generate new ones.

The result? Of the generated structures, 78 percent land within 0.1 electronvolts per atom of the energy minimum after verification. To a layperson that means nothing — to a chemist it means everything: it is a crystal that will not fall apart in your hands. 61 percent of the generated structures are new, that is, absent from any database. The model rediscovered over two thousand real, previously synthesized materials it had never seen in training.

Compared with earlier crystal-structure generators, MatterGen is more than twice as effective at producing structures that are simultaneously stable, unique and novel — and ten times closer to the energy minimum. This is not a cosmetic improvement. It is a leap.

A model that has never seen a test tube designs a magnet without China

MatterGen's most powerful trick is not generating random crystals. It is control.

The model can be fine-tuned to target a specific property. Microsoft's team trained it on 605,000 structures with computed magnetic density and asked it to search for a material with a value of 0.20 Å⁻³ — the threshold where permanent magnets begin. It also trained a version searching for a material with a bulk modulus of 400 gigapascals, that is, superhard. In this second task MatterGen found 106 structures meeting the condition while spending a budget of 180 quantum calculations. Classical screening found 40 within the same budget.

Then something of geopolitical significance happened. The researchers told the model to design a magnet with low supply risk — one that does not require the rare-earth elements controlled by China. MatterGen, asked to combine high magnetic density with a low supply-concentration index, simply eliminated cobalt and gadolinium from its proposals. Without any human hint. It discovered on its own that these two metals were better left alone.

The paradox resolved: a hallucination that passes through the sieve of physics

Let us return to the paradox. How can a machine famous for making things up design materials better than those suggested by human intuition? The answer is not what you would expect: MatterGen never stopped hallucinating. It learned to hallucinate usefully.

The model generates thousands of proposals. Most of them are garbage — just as most sentences a chatbot invents are nonsense. The difference lies in the sieve that filters them. Every structure is subjected to quantum calculations called density functional theory. It is a simulation that predicts where electrons "want" to sit — and whether a given arrangement of atoms will hold together or fall apart.

Imagine a sieve with meshes so fine that only structures lying at the bottom of an energy valley pass through. A material on the slope of the valley is unstable — it will slide down, that is, react into something else. MatterGen throws thousands of "hallucinations" into this sieve, and physics lets through only those that have a right to exist.

And here comes the moment of insight: a hallucination that passes through the sieve of physics ceases to be a hallucination. It becomes a candidate.

TaCr₂O₆: physical evidence

Every claim of miracles demands proof. Microsoft's team decided to deliver it — physically, in a crucible.

The researchers asked the model to design a material with a bulk modulus of 200 gigapascals. That is an enormous value — it means a material that barely changes volume under pressure. The model generated 8,192 candidates. Filters — uniqueness, novelty, quantum stability — narrowed the pool to 75. Experts chose four for synthesis.

One succeeded.

The material with the formula TaCr₂O₆ — a tantalum-chromium oxide — was synthesized in the laboratory. Theory predicted a bulk modulus of 222 gigapascals. The measurement, accounting for an imperfect sample, gave 169 gigapascals. The target was 200. A hit within twenty percent.

The word "hallucination" loses its force here. A material that existed solely as a string of numbers in the model's memory passed into the real world. It was weighed, ground, and measured with an X-ray diffractometer.

As a side note: during this search the model quietly rediscovered 101 materials previously registered in the ICSD crystallographic database that were not in its training data. It did not know them. It simply invented them anew — like a student who independently derives a theorem they were never shown.

A race played by the biggest players

MatterGen is not alone. The race for "generative materials" has become one of the hottest fronts in artificial intelligence.

Google DeepMind announced GNoME — a model that generated 2.2 million new crystal structures and identified 380,000 stable ones. Meta and more than a dozen universities are developing their own generators. The stakes are concrete: a better battery is tens of billions of dollars of market, a cheaper catalyst is a change in the economics of an entire chemical industry, a magnet without neodymium is independence from China.

Microsoft bet on something its competitors did not: experiment. GNoME reported numbers on paper. MatterGen delivered a crystal into your hand. The difference between "the model predicted" and "the laboratory confirmed" is a chasm that investors can measure in dollars.

Microsoft has already shown what it is worth. A year before publishing MatterGen, together with Pacific Northwest National Laboratory, it screened 32 million candidates and found a new battery electrolyte — a material that uses 70 percent less lithium. The search took not months but eighty hours. That was the preview. MatterGen is act two.

MatterGen's code is open. Anyone can download it from GitHub and run it on their own problem. That matters — because it means the race is not reserved for Silicon Valley.

Poland has Grzybowski. It has no MatterGen

Poland has no MatterGen of its own. What it has is Bartosz Grzybowski — a chemist from the Polish Academy of Sciences who a decade ago built Chematica, a program for planning chemical synthesis. Merck bought the license and now sells it as Synthia. Grzybowski, affiliated with the Institute of Organic Chemistry of the PAS and IBS in South Korea, is among the most cited chemists in the world — proof that Polish thinking can stand behind a global breakthrough in AI for chemistry.

AGH in Kraków develops computational methods and machine learning for superconductor research. IDEAS NCBR, founded to build Polish artificial intelligence, funds PhDs at the boundary of AI and the hard sciences. The Warsaw and Wrocław universities of technology train generations of researchers who now work in the labs of Microsoft and Google — abroad.

The paradox is well known: Poland exports talent and imports technology. But MatterGen is open. A Polish PhD student can download the model today, train it on sodium batteries or catalysts — exactly where Polish companies fight for an edge — and start designing materials that exist in no database. The barrier to entry has fallen from a billion dollars to the price of a graphics card.

The money is there too — you just have to reach for it. The National Centre for Research and Development runs the SMART Path, through which companies fund R&D projects; the National Science Centre offers OPUS and MAESTRO grants for teams combining chemistry with machine learning. These are not amounts that would buy Microsoft's supercomputer, but they are enough for the first generative materials with a Polish address.

The choice is simple: in two years a Polish team will either use a generative materials model or compete against a team that uses one. No one wins that second race in the long run.

Steven Schwartz lost his case because he trusted a machine that hallucinates. The chemists in Shenzhen won because they did exactly the same thing — they trusted a machine that hallucinates. The difference? They passed its inventions through the sieve of physics before believing them. That is the skill everyone who wants to work with AI will need: not distinguishing truth from fiction, but knowing when fiction has the right to become truth.

Sources

Zeni C., Pinsler R., Zügner D. i in., A generative model for inorganic materials design, Nature (2025). DOI: 10.1038/s41586-025-08628-5

Merchant A., Batzner S., Schoenholz S.S. i in., Scaling deep learning for materials discovery, Nature (2023). DOI: 10.1038/s41586-023-06735-9

Jain A., Ong S.P., Hautier G. i in., Commentary: The Materials Project, APL Materials (2013). DOI: 10.1063/1.4812323

Xie T., Fu X., Ganea O.-E. i in., Crystal Diffusion Variational Autoencoder for Periodic Material Generation, ICLR (2022). arXiv: 2110.06197