MG3Net will perhaps be the great craftsman of a real revolution in materials science, a bit like AlphaFold for proteins.
The goal of materials scientists is to understand the nature, structure and properties of substances in the world around us. It is an incredibly broad branch of scientific research; it is at the heart of many technical revolutions in excessively important fields, as much for fundamental research as for the daily life of humanity.
But as interesting as it is, it can be quite thankless because of its formidable complexity. Indeed, materials science is at the intersection of many disciplines. To understand matter, you have to explore its chemical, mechanical, thermal and electrical properties, etc. This involves relying on teams of specialists who have advanced skills in many fields.
These are extremely tedious groundwork that tend to produce small incremental advances. Revolutions are therefore quite rare in this area… at least, for the moment. Because in recent years, researchers have followed with interest the rise of a very promising tool in the context of this activity: artificial intelligence.
This technology works wonders when it comes to working with phenomena that are too subtle or complex for the human brain. It has already led to spectacular advances in fundamental research. One can for example quote AlphaFold, the incredible algorithm of DeepMind; it has completely revolutionized the daily life of specialists by offering a huge database of the 3D structure of human proteins (see our file below).
The AlphaFold of materials
Now, materials science researchers hope to achieve a comparable breakthrough in their field through AI. And the idea is extremely attractive, because they could use the same qualities of these algorithms that have already enabled AlphaFold to do its job.
We think in particular of their combinatorial power. In the case of AlphaFold, the AI was able to test billions of combinations and possibilities to determine the famous 3D structures of proteins. And above all, it was able to do it at an incredible speed, far superior to all traditional algorithms based on brute force.
Here, it’s a bit the same idea. On paper, an AI could explore lots of ways to arrange atoms to identify the physical and chemical properties of materials. It could thus make it possible to improve some of them, or even to predict the properties of new substances that are still unknown.
And this is precisely what a team from the University of California at San Diego managed to do with M3GNet, an AI-based tool. Its objective: to build a rather special catalog of materials. They could all exist according to the laws of physics; but they have never been identified and therefore remain hypothetical.
The conceptual proximity with AlphaFold is obvious; so obvious, in fact, that the authors of these works refer to it explicitly in their paper. So we end up with a real Materials AlphaFold Says Shyue Ping Ong, co-author of the study.
31 million unknown materials simulated
As always with these algorithms, it had to be trained first. To do this, they relied on the huge database of the Materials Project. It is a research project whose objective is to build a vast encyclopedia of materials and their properties.
This information was dissected by the program, which then crossed and recombined it in every way possible and imaginable. It’s a bit like dismantling a LEGO construction to assemble new ones from the same parts.
And at the end of this training, like the algorithmic jewel of DeepMind, M3GNet struck a blow. The researchers were able to extract a huge list of more than 31 million hypothetical materials, with their structures and their different properties. And among them, about 1 million would be perfectly stable in theory.
All these results have been compiled in an online database called matterverse.ai. And this is a resource that could be very useful. Most of these hypothetical materials won’t really be of practical use; but this catalog could also house rare pearls with exceptionally interesting properties in certain situations.
A colossal potential for basic research
If researchers need a material that has specific properties, they can use this list to find relatively similar examples, then try to synthesize them. And it’s not just about exploration. Once mature, this approach will likely have very concrete outlets.
Ong cites in particular the example of batteries. “ We are often interested in the rate of diffusion of lithium ions in a Li-ion battery. The faster it is, the faster it can be loaded and unloaded. However, M3GNet can be used to predict the lithium conductivity of a material with good accuracy. “, he explains.
And this is just the tip of a huge iceberg. In theory, this technology could eliminate a significant part of the risky exploration that still handicaps this work. What considerably accelerate research in this field. ” We truly believe that M3GNet is a transformative tool that will greatly expand our ability to explore the chemistry and structure of new materials. concludes Ong.
And the most interesting thing is that this is just the beginning. Ong and his team matter increase massively the number of simulated materialsbut also the number of predicted properties. This will make it possible to focus on those which present a real concrete potential for basic research or industry.
It will therefore be very interesting to follow the evolution of MG3Net. Because if it is not yet quite at the same stage of operational maturity as AlphaFold, it shows just as great potential. In a few years, this tool and its equivalents could well radically change the way materials scientists workwith remarkable advances in countless areas.
The text of the study is available here.
We would like to give thanks to the author of this post for this outstanding web content
An artificial intelligence has identified 31 million unknown materials
Our social media pages here and other related pages herehttps://www.ai-magazine.com/related-pages/