Russian scientists have improved the method of chemical modeling Deep Mind

Scientists from the Russian Quantum Center together with colleagues from NUST MISIS have increased productivity

fermionic neural network (FermiNet), createdsubsidiary of Google, British developer of artificial intelligence systems DeepMind. During the experiment, carried out with the support of the Russian Science Foundation and the Nissan Research Center, specialists used the FermiNet neural network and the cloud quantum computing platform QBoard to simulate larger chemical systems. The results are described in the scientific journal International Journal of Quantum Chemistry. 

Researchers in a variety of scientific fieldsregularly use computing architectures based on artificial neural networks to analyze huge amounts of data and predict the behavior of individual systems. Thus, in 2020, DeepMind for the first time used a fermionic neural network to solve one of the key problems in the field of chemistry - the Schrödinger equation for electrons in molecules. 

Most problems in quantum mechanics cannotbe solved with an accurate answer, so scientists are forced to use approximation - a scientific method that consists in finding approximate values ​​by replacing objects with simplified analogs. By varying the free parameters, physicists manage to find wave functions that most accurately describe the state of the system. This form of search - ansatz - is actively used in quantum chemistry, since the modeling of elementary chemical reactions is still given to scientists with great difficulty, even for a small number of atoms in a system.

As part of the experiment, a joint team ofphysicists, chemists and machine learning specialists used the FermiNet architecture as an ansatz. Next, the experts began to iteratively improve the neural network through an updated procedure for training it. During the calculations, tools from the cloud quantum computing platform QBoard were used. Scientists have not only been able to simulate higher-dimensional systems than the original FermiNet architecture allowed, but also have increased the accuracy of classical calculations in electron-nuclear and electron-electron interactions. 

The results have been demonstrated in the processmodeling of nitrogen, carbon monoxide, ethylene, hydrogen fluoride and a number of other molecules. In the future, the data obtained can be used in pharmacology to create new drugs, materials science and the fuel industry.

“A combination of machine learning methods andquantum chemistry today gives very interesting results. Such interdisciplinary interactions of physicists, chemists, biologists, programmers lead to the enrichment of classical approaches and such interesting hybrid solutions as our case on using QBoard to develop the FermiNet network,” said Alexey Fedorov, head of the Quantum Information Technologies research group at the Russian Quantum Center.

Read more:

Named a vitamin that protects the brain from dementia

See how the Earth has changed over 100 million years on the most detailed map

It turned out which men are the most fertile: their sperm is 50% better than the rest