Giovanni Pizzi and his team at the Paul Scherrer Institute in Villigen, Switzerland, have cracked a puzzle that has quietly blocked thousands of material simulations: they taught an artificial intelligence to see the invisible hydrogen atoms hidden inside crystal structures.
The problem sounds deceptively simple. Hydrogen is everywhere in crystalline materials—it binds atoms together, alters electrical properties, enables superconductivity, and determines how well a material might store energy. Yet traditional X-ray diffraction methods, the workhorses of materials science, struggle to detect hydrogen atoms precisely. Their positions vanish from the databases and visualizations that scientists rely on. Without knowing where every atom sits, you cannot simulate how a material will behave. As Pizzi explains, this leaves researchers unable to work with "several thousand potentially interesting materials" for computation. The consequence is innovation delayed or possibilities never explored.
The solution came from an unexpected direction: computer vision. Pizzi's team, working with collaborators from the universities of Parma and Modena in Italy, adapted the diffusion model techniques that AI systems use to fill in missing elements in photographs—imagine reconstructing a hidden dog's paw in a picture. They developed an open-source tool called XtalPaint, built on Microsoft's MatterGen model, that applies this logic to atomic structures. Instead of adding noise to an entire crystal and reconstructing it blindly, XtalPaint adds noise only to the regions where hydrogen atoms are missing, letting the known atomic positions guide the reconstruction from start to finish. This targeted approach is more efficient and far more accurate than earlier methods.
The results are striking. When the researchers deliberately removed hydrogen positions from known crystal structures and used XtalPaint to find them again, the AI succeeded with a 97% success rate—landing on the correct positions in 87% of cases and, remarkably, finding even more energetically stable configurations in another 10%. That combination of accuracy and discovery power opens a new frontier. Thousands of experimentally confirmed but theoretically inaccessible materials—those hydrogen-rich compounds locked out of simulation work—can now be studied for the first time. The implications reach across materials science: better catalysts, more efficient hydrogen storage systems, potentially even new superconductors.
"If the information about the hydrogen atoms is missing, that's a problem," Pizzi said, understating the scope of what XtalPaint now solves. The method is not limited to hydrogen either. The team has shown it can locate missing atoms of other light elements like lithium and sodium, expanding its utility further.
What makes this story resonate is its elegance: a bottleneck created by the limits of experimental detection, solved by borrowing the language of image restoration. The researchers have made XtalPaint open-source, inviting the global community to apply it to their own materials questions. The work, published in npj Computational Materials, represents the kind of foundational advance that quietly multiplies discovery downstream. Materials scientists can now simulate thousands of compounds they could not reach before, and with each simulation, the possibility of finding the next transformative material—one that stores hydrogen more efficiently, conducts electricity without loss, or unlocks some property we haven't yet imagined—grows closer to real.