Neural network developed by UNN scientists can predict performance of materials

A neural network capable of predicting the performance of materials for photonic technologies has been developed at Lobachevsky University. This model can estimate the level of "topological protection" of photonic crystals with an accuracy of more than 90%, thus providing new opportunities for developing and enhancing the performance of computing systems.

According to the researchers, transmitting information using light particles (photons) makes it possible to extend the range of signal transmission and make the devices themselves more compact. The potential of such photonic technologies can be realised using modern methods of optical radiation control. Photonic crystal lattices of light waveguides allow photons to be controlled and directed. However, light can be scattered by defects, which are often introduced during the manufacturing process. This leads to distorted signals and loss of information. To protect the crystal from negative effects, "topological" systems with a special structure are used.

"The lattice's special symmetry can protect light in a photonic system from strong scattering. In this case, information is transmitted along selected boundaries using so-called edge states of the light. Using data on how radiation passes through photonic crystals, we trained a neural network to analyse the structural features of the sample and predict its ability to maintain edge states," said Lev Smirnov, author of the study and head of the Artificial Intelligence and Processing of Large Data Arrays research laboratory at the UNN Institute of Information Technology, Mathematics and Mechanics (IITMM).

Previously, additional experiments, measurements and mathematical calculations were required to determine whether a crystal was protected from light scattering. The approach of Nizhny Novgorod scientists makes it possible to draw a conclusion on the basis of a single measurement of the basic parameter of the signal intensity at the lattice output.

"To determine the topological properties of an optical element using our neural network, experimenters and engineers only need to measure the signal intensity in the central region of the sample and feed this data into the trained model. This method saves considerable time and simplifies computation. Such optimisation is particularly relevant today as photonic elements are integrated into classical and quantum electronics. In the future, they could serve as the basis for photonic computers," says Ekaterina Smolina, junior researcher of the Artificial Intelligence and Processing of Large Data Arrays research laboratory at IITMM.

The results of this research were published in the journal Nanophotonics in 2024.