DeepMind releases AI tool GNoME, said to have successfully discovered 2.2 million new crystal materials

According to news on December 1, Google's DeepMind recently demonstrated its own AI tool GNoME in the journal "Nature" and introduced the related applications of AI in materials science. It is reported that DeepMind used GNoME to discover 2.2 million 380,000 new crystals, of which 380,000 are stable materials that can be made in laboratories and are expected to be used in batteries or superconductors.

The content that needs to be rewritten is: ▲Image source deep learning

Currently, in the ICSD data, about 20,000 crystals are considered to be computationally "Steady state", Materials Project and other teams have previously identified 28,000 more crystals through a series of calculation methods. However, DeepMind believes that although the previously improved calculation methods in the industry can speed up the discovery of new crystal structures, the cost of time and money is quite high.

DeepMind’s new tool GNoME is said to have broken through various previous calculation methods and can accurately predict a series of stable crystal structures and generate 2.2 million materials from them. DeepMind claims that

If It would take 800 years to calculate these materials by human power alone.

The content that needs to be rewritten is: ▲Picture source deep learning

This site learned from the DeepMind report that the efficiency of GNoME development materials is quite high. The model designed a total of 52,000 new graphene layered compounds, whereas before, humans had only identified about 1,000 similar materials. Additionally, GNoME discovered 528 potential lithium-ion conductors with up to 25 times the conductivity of previous materials. Scientists believe that the above-mentioned discoveries alone are expected to improve the energy consumption of batteries currently used in electronic products.

The content that needs to be rewritten is: ▲ Deep learning of the picture source

DeepMind pointed out that GNoME uses two strategies to search for materials, one is based on existing The other is based on chemical companies and explores candidate structures in a more random way. The model uses neural networks to simultaneously process and analyze the output of both methods, and uses Density Functional Theory calculations to evaluate the stability of these material candidates. In addition, GNoME also uses the "Active Learning" method to improve the accuracy and efficiency of crystal prediction, thereby significantly increasing the speed and success rate of discovering new materials

The content that needs to be rewritten is: ▲Image source deep learning

The GNoME model aims to reduce the cost of discovering new materials.

Currently, scientists around the world have produced 736 types of GNoME predictions in the laboratory New materials, which proves the accuracy and feasibility of GNoME's crystal predictions in reality, and DeepMind has now made GNoME's newly discovered crystal database public to assist researchers in testing and manufacturing candidate materials.

The above is the detailed content of DeepMind releases AI tool GNoME, said to have successfully discovered 2.2 million new crystal materials. For more information, please follow other related articles on the PHP Chinese website!