A chunk of nuclear fuel inside a reactor is not just a solid. When heated, the uranium atoms in it start to 'wiggle' their magnetic moments, and this motion changes thermal conductivity and other properties. Previously, accurately calculating such behavior was nearly impossible: quantum computations demand unimaginable resources, and conventional models ignore the coupling between magnetism and crystal lattice vibrations.
Now a neural network has come to the rescue. It was fed thousands of quantum calculations for different magnetic configurations of uranium, and it learned to predict the energy and forces acting on the atoms. With this neural network, scientists modeled the behavior of millions of atoms and saw how magnetic order breaks down upon heating, like a row of soldiers that dissolves into a crowd.
The model still errs in details: at low temperatures it predicts a different magnetic state than what is observed in experiments. But it reproduces the transition itself with the correct temperature. This is an important step: it means neural networks can replace expensive quantum calculations for large systems and help design safer and more efficient nuclear fuel.
Uranium dioxide melts at 2865 °C — hotter than the surface of molten lava but cooler than an electric welding arc.