From new energy vehicle batteries to solar cells, to computer chips and many other fields, once new materials are discovered, it undoubtedly accelerates technological breakthroughs. However, the development of new materials typically requires scientists to spend months or even years conducting repeated experiments and verifications. However, a study recently released by Google's DeepMind may greatly accelerate the application speed of new materials in many technological fields.
The research team of Google DeepMind has discovered up to 2.2 million theoretically stable but largely unrealized crystal structures through the artificial intelligence tool "Material Exploration Graph Network" (GNoME). This achievement was published in the top issue of Nature on November 29th.
The number of crystal structures discovered by GNoME is more than 45 times that of such substances discovered in scientific history. The industry believes that this technology provides a new path for the development of renewable energy and advanced computing chips.
GNoME stability prediction accuracy rapidly improves in iterative learning
It is reported that this artificial intelligence model, known as GNoME, aims to predict the inorganic crystal structure, that is, the repeated arrangement of atoms, to give a certain material special properties. So far, only about 48000 types of inorganic crystals are known to humans.
The GNoME model extends this number to up to 2.2 million different types. Deepmind stated that among these 2.2 million new crystal structures, 380000 stable crystal structures are expected to be synthesized through experiments, which have practical application prospects and may develop "revolutionary technologies for the future," such as superconducting materials and next-generation battery materials. GNoME "has achieved an order of magnitude expansion in stable materials known to humans, discovering new materials with revolutionary potential for approximately 800 years."
In order to discover more new materials, the DeepMind team combined two different deep learning models. The first approach is to modify the elements in existing materials, resulting in over 1 billion structures. The second method sets aside existing material structures and predicts the stability of new materials solely based on chemical formulas. The combination of these two deep learning models provides a wider range of possibilities for the discovery of new materials.
After the generation of candidate new material structures, researchers screened them using the GNoME model. This model can predict the decomposition energy of a specific structure, which is an important indicator for measuring the stability of materials. Only materials that are stable and not easily decomposed can have significant implications for industrial use. Therefore, GNoME will predict and select the most promising materials for application, and further evaluate them based on known theoretical frameworks.
It is reported that the above process will be repeated multiple times by the DeepMind team, and each discovery will be incorporated into the next training. Although in the first round of testing, the accuracy of GNoME in predicting the stability of different materials was only about 5%, the prediction accuracy of GNoME rapidly improved throughout the entire iterative learning process. The final results indicate that the accuracy of GNoME in predicting structural stability in the first model has exceeded 80%, while in the second model, the accuracy has improved to 33%.
Although some of the new structures may decay into more stable forms or may not be fully created, the DeepMind team has successfully created 736 new materials discovered by GNoME in the laboratory, including alkaline earth diamond like optical materials (Li4MgGe2S7) and potential superconductors (Mo5GeB2), which are currently being tested.
"For me, materials science is basically the intersection of abstract thinking and the physical universe, and it's hard to imagine a technology that won't be improved by better materials," said Dougus Cubuk, one of the co authors of DeepMind's aforementioned paper and head of materials research and development
Ju Li, a professor of materials science and engineering at the Massachusetts Institute of Technology, believes that GNoME can be seen as an "alpha fold" in the field of material discovery. Alpha Fold is an artificial intelligence system launched by DeepMind in 2020, which can accurately predict protein structures and has made significant progress in biological research and drug discovery. Ju Li said that thanks to the powerful capabilities of GNoME, the number of known stable materials in humans has increased nearly tenfold, reaching 421000 species.
GNoME has discovered over 500 promising lithium-ion conductors with potential applications
The Daily Economic News reporter noticed that using artificial intelligence models to manufacture new materials was not the first of DeepMind's kind - the Materials Project, led by Kristin Persson from Lawrence Berkeley National Laboratory in the United States, has used similar techniques to discover and improve the stability of 48000 materials. This experiment obtained data from a material database, including some discoveries of GNoME, and designed new materials using machine learning and robotic arms without human intervention.
However, the new materials discovered by GNoME set it apart from the work of Lawrence Berkeley National Laboratory in terms of scale and accuracy.
Chris Bartel, assistant professor of chemical engineering and materials science at the University of Minnesota, believes that GNoME has at least one order of magnitude more training data than any previous model. Yifei Mo, Associate Professor of Materials Science and Engineering at the University of Maryland, also pointed out that conducting similar research in the past was not only costly but also limited in scale. GNoME can expand the discovery of these new materials with higher accuracy and lower computational costs, "and the impact may be enormous."
More importantly, the DeepMind team has collaborated with the Berkeley National Laboratory and created a robotic laboratory called A-Lab that can autonomously synthesize these new crystals. After the discovery of new materials, it is equally important to synthesize and validate their applications. A-Lab has also been combining some of the discoveries of GNoME with its "material project" achievements, that is, combining robotics technology with machine learning to optimize the subsequent development of these materials.
Researchers from DeepMind and Berkeley Labs suggest that these new artificial intelligence tools can help accelerate hardware innovation in energy, computing, and many other fields. For example, lithium-ion battery conductors are one of the most promising cases of new materials discovered by GNoME artificial intelligence models. DeepMind stated that GNoME has discovered 528 promising lithium-ion conductors, some of which may help improve the efficiency of electric vehicle batteries.
However, even after the discovery of new materials, it usually takes decades to push them towards commercial applications. Dogus Cubuk said at a press conference, "If we can shorten the process from discovery to application to 5 years, it will be a great progress."