AI speeds up the search for materials to store nuclear waste safely

# AI speeds up the search for materials to store nuclear waste safely

Russian scientists have found a way to tackle one of the hardest problems in nuclear power: finding materials that can store nuclear waste reliably. The key is an unusual pairing of classical chemistry and machine learning. A team of researchers from Skoltech, AIRI and Sber AI built an accurate phase diagram of technetium carbides using graph neural networks instead of traditional quantum-mechanical calculations. The results were published in the authoritative journal Acta Materialia. This is more than a scientific result: it shows how AI can radically change the pace of development in materials science.

The task was genuinely difficult. Technetium carbides are critical for the safe disposal of nuclear waste, but studying their properties demands enormous computing resources. The traditional approach relies on quantum-mechanical calculations, methods that model the behavior of electrons and nuclei from the underlying laws of physics. These calculations are extremely accurate, but they take days or weeks even on powerful supercomputers. When thousands of candidate materials and their combinations need to be examined, the task becomes practically impossible. Scientists hit a dead end: without a full understanding of a material's structure, it is impossible to guarantee its reliability for storing hazardous substances, and calculating that structure takes far too long.

This is where graph neural networks come in. These models represent molecules as graphs in which atoms are nodes and chemical bonds are edges. The architecture lets the network grasp spatial relationships between atoms and predict their behavior. The researchers trained the network on the results of classical quantum-mechanical calculations, letting it "learn" the patterns those calculations had revealed. Once trained, the model could predict the properties of new compounds almost instantly, something that previously would have taken weeks. It is like training an experienced craftsman and then asking him to write a summary of his knowledge: other people can then use the summary instead of going through a long apprenticeship.

The result is impressive. Thanks to the AI model, the researchers built a detailed phase diagram of technetium carbides, a map showing which crystal structures of the material are stable at different temperatures and pressures. The process became thousands of times faster without losing accuracy. This means scientists were able to study in a few weeks a volume of materials that would previously have taken years. The work shows that machine learning in materials science is not a demonstration toy but a full-fledged tool for real scientific problems.

The significance of this advance goes well beyond technetium. The nuclear industry needs materials that can withstand extreme conditions: high temperatures, radiation and corrosion. All of this requires a deep understanding of material structure, and traditional methods are slow. Applying AI opens the way to a systematic search for new compounds with predefined properties. This may speed up the development of safer waste disposal systems, and possibly of new generations of nuclear reactors.

The demonstration of neural networks handling chemistry and materials science signals deep changes in science. AI does not replace scientists, but it is no longer just a data analysis tool. Machine learning is becoming a partner in the creative process of finding solutions. By shortening the path from theory to practice, such approaches accelerate not only research but the very pace of scientific progress in fields where every day counts.