A photocatalyst is a powder that, when placed in water under light, splits it into bubbles of hydrogen and oxygen. The most popular material for this is titanium dioxide, the white pigment in paints and toothpaste. But a pure crystal works sluggishly: it absorbs only ultraviolet light, and such rays are scarce in the solar spectrum. So the crystal is deliberately spoiled — defects are introduced: individual oxygen atoms are removed, or part of the titanium is replaced with trivalent ions. This "dirty" powder turns gray or even black and starts using visible light.
The problem is that there are hundreds of recipes for this spoiling. Some heat the sample in hydrogen for an hour at 400 °C, others for two hours at 600 °C, and still others irradiate it with plasma. The results are scattered across dozens of papers, and comparing them is difficult: authors measure activity differently and don't always report important details. Conventional machine analysis stalls here — the data are too heterogeneous. The authors took a different approach: they manually selected publications that describe synthesis conditions and final hydrogen production for defective titanium dioxide and turned them into a unified database.
Then the researchers trained a large language model not just to memorize patterns but to reason based on rules extracted from the papers. The model doesn't say, "In my experience, 500 °C is good." It finds publications in the database that compared different temperatures and links them to the mechanism: at too low a temperature, there are few defects; at too high a temperature, they coalesce and stop helping. This approach is called retrieval-augmented reasoning: the model relies on specific sentences from papers, like a student using cheat sheets, so its conclusions can be verified.
The practical benefit of such an AI assistant is not just faster recipe discovery. It highlights gaps: if there is little data for a certain condition, the model honestly says confidence is low. This approach can be transferred to other materials and other reactions — anywhere science has accumulated mountains of scattered but valuable observations.
Titanium dioxide is the very pigment that makes paint, paper, and even the glaze on donuts white. By adding a bit of "dirt," it's turned into a devourer of sunlight.