Google DeepMind discovers new materials via deep learning

AI robot is performing material synthesis. [Courtesy of Lawrence Berkeley National Laboratory]

DeepMind, the artificial intelligence (AI) development company under Google LLC’s Alphabet Inc., announced on Wednesday local time that it identified more than 2.2 million crystals that could revolutionize technology.

DeepMind managed to predict the structures, components, and characteristics of the materials via an AI analysis of all material substances created by humans, such as electronic devices and airplanes. This achievement is equivalent to the amount of knowledge that can be accumulated by humanity over 800 years.

Industry insiders noted that the use of predictive results could lead to advances in the field of materials science, enabling developments such as achieving superconductivity or significantly improving battery efficiency.

A team of researchers at DeepMind, led by Dr. Amil Merchant, introduced Graph Networks for Materials Exploration (GNoME), a new deep learning tool that analyzes materials, in the international journal Nature.

GNoME has learned from all materials discovered by humanity until now, comprehensively studying the structures, components, and characteristics of over 48,000 materials.

The research team based GNoME’s learning on active learning, where AI selects data that seems most effective for learning from the available data.

This is contrary to passive machine learning where humans label the data for AI learning.

Through this learning method, GNoME identifies patterns that go beyond the original training data.

GNoME specializes in finding hidden relationships between materials based on data. It recombines the elemental ratios within materials, compares the performance changes with ratio variations, and identifies materials with optimal performance for specific purposes.

The success rate of predictions exceeds 80 percent, significantly surpassing the previous success rate of 50 percent achieved by conventional analysis methods.

GNoME has discovered about 2.2 million new materials and of these, analysis indicates that around 381,000 are stable materials suitable for actual use. The research team suggests that about 52,000 of these materials can be used in batteries.

It also identified around 528 candidates for lithium-ion conductors that could significantly enhance battery performance and efficiency.

The remaining task is to synthesize the discovered materials into actual substances. Typically, synthesizing a material involves a trial-and-error process that takes several months or even years.

The problem has been addressed by a research team at the Lawrence Berkeley National Laboratory in the United States.

Material scientists have also publicly revealed results in Nature, demonstrating that AI robots can quickly perform material synthesis using the learned AI synthesis method.

The AI, trained in material synthesis methods, identifies the optimal synthesis method and the robot then carries out the synthesis.

In practice, this AI robot, utilizing the analysis results of the Google DeepMind research team, successfully synthesized 41 materials in just 17 days.

DeepMind plans to release the data on the discovered 381,000 materials. While making the AI code public, it also plans to gradually release all the data obtained from the discovery of 2.2 million materials.

“Developing new materials means finding the optimal combination by changing the elemental ratios,” according to Han Sang-soo, head of the Computational Science Research Center at the Korea Institute of Science and Technology. “If AI takes over this part, it can dramatically reduce the time required for developing new materials.”