Russian scientists have developed a neural network system capable of calculating the behavior of superconductors at the microscopic level. The results of the work were shared by the Center for Scientific Communication of MIPT.
Superconductors are materials that, under certain conditions, completely lose electrical resistance. They are considered a promising basis for the energy sector of the future, medical diagnostics, and high-precision electronics. However, their study is associated with extremely complex mathematics and large computational costs.
Today, to describe the behavior of particles inside a superconductor, physicists use the Bogolyubov–de Gennes equations. The solution of these equations limits the scale of models: usually, samples with a size of only a few hundred atoms are studied. This makes it especially difficult to analyze real materials with randomly distributed defects.
Specialists from MIPT, together with colleagues from HSE University and NRNU MEPhI, proposed an alternative approach. They trained a neural network on data obtained from accurate calculations of small sections of a superconductor with a size of 24×24 atoms. After training, the algorithm learned to predict the properties of significantly larger structures — over 100×100 atoms.
The development allows us to see how "islands" of superconductivity are formed, how they interact with each other, and under what conditions the material loses its conductive properties. According to researchers, the technology can significantly speed up the search for new stable superconductors and simplify the modeling of quantum transitions.