"Deep thinking" has a new breakthrough
Neural network opens a new window for understanding electronic interaction
Science and Technology Daily, Beijing, December 9th (Reporter Zhang Mengran) According to a paper published in Science on the 9th, a new study by the famous artificial intelligence company "Deep Thinking" shows that neural networks can be used to build more accurate electronics than before. Density and interaction diagram.
The research helps scientists better understand the interactions between electrons that hold molecules together. It also shows the promise of deep learning to accurately simulate matter at the level of quantum mechanics, enabling researchers to improve computer design at the nanometer level. Explore issues related to materials, drugs, and catalysts horizontally.
Density functional theory (DFT), which describes the basic properties of quantum matter, was first established more than 50 years ago and has become the main method for predicting the characteristics of electronic interactions in chemistry, biology, and materials science.
However, the exact nature of the mapping between electron density and interaction energy, the so-called density functional, has not been understood for a long time.
Therefore, even the most advanced DFT functionals are plagued by basic systematic errors when describing the fractional electron charge and spin.
To address these limitations, "Deep Thinking" researcher James Kirkpatrick and his colleagues used the company's platform to develop a framework for training neural networks on accurate chemical data and fractional electronic constraints to produce functional "DM21".
By expressing the function as a neural network and incorporating precise attributes into the training data, DM21 is able to learn the function without two important system errors (delocalization error and spin symmetry breaking), which results in better performance than previous platforms Simulate a wide range of chemical reaction categories.
The importance of DM21 is not that it produces the final density functional, but that the artificial intelligence method solves the fractional electron and spin problems, which hinder the creation of direct analytical solutions for functionals.
In the short term, this will allow researchers to improve the approximation of the precise density functional through code availability for immediate use.
In the long run, this further shows the prospect of deep learning to accurately simulate matter at the quantum mechanics level. It will allow researchers to explore issues related to materials, drugs, and catalysts at the nanoscale level to achieve material design in computers.
"Understanding nanotechnology is becoming more and more important to help us deal with some of the major challenges of the 21st century, from clean electricity to plastic pollution." Kirkpatrick said, "This research is a step in the right direction. We can better understand the interactions between electrons, which are the glue that holds molecules together."
In order to accelerate the progress of this field, "Deep Thinking" has published this paper and provided open source code for free.