Artificial intelligence has learned to determine the properties of any molecules by solving the Schrödinger equation

Artificial intelligence has learned to determine the properties of any molecules by solving the Schrödinger equation
Artificial intelligence has learned to determine the properties of any molecules by solving the Schrödinger equation
Anonim

A breakthrough algorithm can solve the Schrödinger equation for arbitrary molecules within a reasonable time frame and without involving supercomputers. This makes it possible, without laborious and costly field experiments, to determine the basic properties of a substance with a high probability.

Artificial intelligence learned to determine the properties of any molecules by solving the Schrödinger equation

The development was presented by German scientists from the Free University of Berlin (Freie Universität Berlin). They described the specifics of developing and training a deep neural network PauliNet in an article that was published in the peer-reviewed journal Nature Chemistry. Since the material is available only by subscription, its main theses can also be studied in the preprint posted on the arXiv portal a year earlier. Since then, scientific work has been substantially supplemented, including practical results, but a general idea of ​​the technology is provided by a preliminary publication.

The PauliNet algorithm is named after the Pauli principle, one of the fundamental rules of quantum mechanics. According to this principle, two or more electrons in atoms cannot be in the same quantum states. That is, during the exchange of electrons, their wave function changes sign. This antisymmetry, as well as a number of other postulates of quantum physics, were "hardwired" into the Deep neural network at once. But she was already taught other properties of elementary particles - in particular, the complex laws of the distribution of electrons over the shells around the nuclei of atoms.

Based on these data, the neural network learned to explore arbitrary molecules using quantum Monte Carlo methods. They involve solving the Schrödinger equations for a large number of particles. The main difficulty in performing such tasks is the need for large computing power to determine the multiparticle wave function. Usually, simpler methods are used, such as Density Functional Theory (DFT) or Coupled Clusters (CC).

However, such simplifications create a number of limitations and are practically useless for many connections. As a result, physicists and chemists have to constantly look for compromises: either low accuracy, but relatively fast calculations, or high accuracy, but at the same time it is necessary to look for what "hardware" all this can be calculated on. And in most cases, there is no particular choice: complex molecules are too tough even for modern supercomputers and distributed computing systems.

And the PauliNet neural network managed to create its own method for calculating wave functions. This algorithm is capable of solving the Schrödinger equations for almost any molecule within a reasonable time frame.

In the given examples, the authors of this artificial intelligence determined the properties of a number of compounds in a matter of tens of hours of operation of ordinary graphic cards of personal computers. Thus, German scientists have found a new and extremely efficient way to calculate the ground state of arbitrary molecules.

Popular by topic