MIPT has accelerated the search for new materials with specific optical properties.

The results have been published in the journal Science China Materials. Materials with a high birefringence index (Δn), particularly in the deep ultraviolet range, are actively used in the development of polarization and various other devices. Traditional experimental methods for discovering such materials are time-consuming and often complex to implement. Therefore, scientists have developed a straightforward and comprehensible method for selecting materials with a high Δn index, utilizing quantum-mechanical models and modern machine learning technologies.

When training a neural network, it is crucial to correctly select the features that describe the training dataset. The optical properties of a material depend on both its crystalline structure and chemical composition. Thus, the model architecture must identify the relationships between chemical composition, crystalline structure, and the material's target properties.

One such architecture that combines the aforementioned properties is a graph neural network. In this study, the ALIGNN graph neural network model was used as the primary tool. Researchers trained it to predict the birefringence of the material based on crystalline structures. To determine the accuracy of predictions, physicists added a D-optimality criterion. This criterion indicates how reliable the model's predictions are on new data and evaluates how well the characteristics of new materials "fit" into the dataset used for model training.

“The proposed approach, which includes collecting a training dataset, developing machine learning models, and screening databases, allowed us to discover new materials with specified optical properties. In the future, this approach may be expanded to other classes of materials,” says Ivan Kruglov, head of the Material Computer Design Laboratory at MIPT.

“Previously, we had not seen works where the D-optimality criterion was used not only to assess the reliability of machine learning model predictions but also to evaluate the 'novelty' of the data. Additionally, the developed approach combines machine learning methods with traditional quantum-mechanical calculations, ensuring accuracy and good interpretability of results. We continue to develop models using new approaches, and perhaps their quality will improve even further in the near future!” explains Lyudmila Bereznikova, a researcher at the Material Computer Design Laboratory at MIPT.

Using the trained graph neural network, after screening the Materials Project database, scientists discovered two new materials with high birefringence values in the deep ultraviolet region: NaYCO₃F₂ and SClO₂F. The discovered materials exhibit Δn = 0.202 and Δn = 0.101 at a wavelength of 1064 nm. Analysis revealed that the high refractive index in NaYCO₃F₂ is attributed to the arrangement of planar structural elements [CO₃]. In the case of SClO₂F, the anisotropic S–Cl bonds make a significant contribution to the Δn value.

This work demonstrates for the first time the possibility of simultaneously using a graph neural network in conjunction with the D-optimality criterion for accurate Δn prediction and the search for new optical materials. The developed model is presumed to be adaptable for finding materials with other physical properties, such as hardness or thermal conductivity.

The research involved scientists from MIPT, the Xinjiang Technical Institute of Physics and Chemistry of the Chinese Academy of Sciences, and the XPANCEO New Technologies Research Center (UAE).