A.I. for research with GNoME: Graph Networks for Material Exploration

In March 2016, DeepMind's AlphaGo A.I. unequivocally defeated the then Go champion, Korea's Lee Sedol. The question of how an A.I. expert in Go would ever be useful was not officially on the agenda, but the expertise developed would not remain untapped for long.

For those unfamiliar with the principles of the game of Go, each player tries to surround his opponent's pieces. As each player places a piece in turn, the way in which the pieces are arranged represents the essence of the strategy, as each shape drawn by the pieces has different properties: a line, a T, a U all have different advantages depending on location and proximity.

In the world of atoms and molecules, the constituents of a material may be the same, but their arrangement and inter-relationships result in different properties. While the game of Go takes place on a 2-dimensional plane with two different elements, the material world is three-dimensional and is made up of almost 100 different atoms and millions of molecules, each with its own mass, charge, shape and characteristics, resulting in particular properties, affinities and behaviors unique to each combination.

GNoME A.I. to the rescue of chemistry

Stable compounds make up most of the matter we know.

Sodium chloride (NaCl - the salt), for example, is a compound that crystallizes easily and dissolves just as easily, a common compound under our conditions. Compounds as simple as carbon can be organized in many different ways to give materials as common as graphite, as rare as diamond or as sophisticated as graphene or fullerenes. Complex combinations, such as those at the basis of cement (silicic acid, alumina, iron oxide and lime), are all composed of stable molecules whose interactions determine the properties of the final product.

In short, of the billions of billions of ways in which atoms and molecules can be combined, only a certain number can produce stable compounds, and of these, even fewer possess specific characteristics such as conductivity, elasticity, magnetism, solidity, strength and so on.

The challenge in materials research is first to determine whether a combination of atoms will be stable, and then to try to predict its properties.

Using data on the 40,000 known materials, and existing theories such as density functional theory (DFT), GNoME, an artificial intelligence developed by DeepMind, the same people behind AlphaGo and AlphaFold, the protein database, trained itself and then modeled over 2.2 million potential crystals, which at the speed of research to date would have taken around 600 years to achieve.

Of these 2.2 million crystals, this A.I. identified that around 421,000 of them would be stable enough to be synthesized, and 738 of them actually were, with properties in agreement with predictions of the order of 80%. Compared with 33% in the best cases to date, or 1% not so long ago, this is unprecedented and opens up destabilizing prospects. A.I. does not yet take into account the effects of these hypothetical compounds on the environment, health or life. Will it eventually provide the formulas for their recycling or recovery?

At present, A.I. is limited to compounds with no more than 5 types of atoms, and to strategies for substitution, estimating bond energy levels and structural forms. As learning progresses, more and more data is accumulated, and other, bolder approaches will radically increase the number of possibilities. This will keep the industry busy for many years to come.

Return of the Go player

In the history of A.I. and the game of Go, in 2023, Kellin Pelrine, a PhD student in Machine Learning at McGill University, soundly beat AlphaGo 14 games to 1 simply by studying the way AlphaGo worked and using a strategy of confusion (recommended by another A.I.), a strategy that any human would have immediately detected.

GNoME's limitations are of the same kind, and its potential only improves with the human interventions that provide it with theories, data and goals to pursue. It's an extraordinary tool whose limits we can recognize... and overcome.

References

Scaling deep learning for materials discovery
https://www.nature.com/articles/s41586-023-06735-9

Millions of new materials discovered with deep learning - Amil Merchant and Ekin Dogus Cubuk - DeepMind
https://deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

AphaFold - https://alphafold.ebi.ac.uk/

Graphene - https://fr.wikipedia.org/wiki/Graph%C3%A8ne

Kellin Pelrine - https://www.linkedin.com/posts/kellin-pelrine_kellin-pelrine-how-he-crushed-a-superhuman-activity-7073179974484021248-hr_F/

Artificial intelligence loses go game to human (helped by AI)
https://www.20minutes.fr/high-tech/4024717-20230221-intelligence-artificielle-perdu-jeu-go-face-humain-aide-ia

Art of Go - Go game rules - https://artdugo.fr/regles-du-jeu-de-go/