Chemists from Russia taught a computer to predict the properties of elements

© Press-service Mptotic the artist imagined inter-atomic bonding in the crystalChemists from Russia taught a computer to predict the properties of elements© Press service of MIPT

. Russian scientists have created a machine learning system that can accurately predict the behavior of the atoms of aluminium and uranium at different temperatures and pressures, which will allow in the future to create new materials of these metals, using a computer, according to an article published in the journal Scientific Reports.

«The magnitude of the forces of intermolecular interactions of atoms in crystals can be successfully applied to predict the behavior of atoms of that element at other temperatures and in other phase States. Showing the state of a substance depending on pressure and temperature, phase diagrams allow to determine the possibilities and limits of the use of elements», says Ivan Kruglov of Moscow physical-technical Institute, quoted in a press-service of the University.

Many chemists and physicists for over a century the dream of acquiring the ability to predict the properties of arbitrary materials and chemicals, knowing only the physical properties of individual atoms. The realization of this dream will allow scientists to create materials with precisely defined properties that are not trying to find them «blind» that is going to revolutionize medicine, engineering, construction and many other areas of life and industry.

Artem Oganov, Professor of Skoltech and head of laboratory at MIPT, for several years working to realize this dream by creating different computer models and tools to predict the properties of various crystals and certain chemical substances in unusual conditions, for example in the cores of planets or at high pressures.

Modern programs allowing such calculations, carried out actually on the «low», quantum level, assessing how changing the length and strength of chemical bonds, their energy, the arrangement of atoms relative to each other, their charge and other fundamental physical parameters.

This technique works very well, but it requires a huge amount of computing resources, making it extremely difficult to use for studying and predicting the properties of large molecules, including proteins or potential drugs. Other options allow you to calculate as accurately, scientists until recently were not, which made the prediction of the properties of different compounds, alloys, and complex molecules is extremely difficult thing.

Oganov and his colleagues at the life Sciences research Institute of automatics named after N. L. Dukhova found a way to speed up these calculations, with the involvement of machine learning systems. As suggested by scientists, computers can learn to predict how atoms interact with each other, even in complex molecules, studying simple examples, calculated by using the classical methods of quantum mechanics. This allows you to get rid of some complex steps such as miscalculation energy, which usually spend the lion’s share of time.

The team Oganova checked by calculating the properties of aluminium and uranium at different temperatures and pressures. Aluminium, as the researchers note, has been well studied in the past, and the properties of uranium, on the contrary, still a matter of controversy among physicists.

Comparing the results of machine calculations and reference data on the properties of these metals, scientists confirmed that their methods of predicting properties of chemicals works as well as quantum mechanical calculations, but it requires thousands of times less computing resources.

The method of calculation of properties of materials, as the researchers note, can be adapted to work with other chemical elements, which significantly enhance its practical applicability.