In March 2016, DeepMind's AlphaGo AI handily beat the then Go game champion, Korean Lee Sedol. The question of how an AI expert in Go might one day be useful was not officially on the agenda, but the expertise developed should not go unused for long.
For those who do not know the principles of the game of Go, remember that each player tries to surround his opponent's pawns. As everyone takes turns placing a figure, the way the figures are arranged represents the core of the strategy, as each shape drawn by the figures results in different properties. A line, a T, a U have different advantages depending on the location and neighborhood.
In the world of atoms and molecules, the components of a material may be the same, but their arrangement and interrelationships result in different properties. If the game of Go takes place on a two-dimensional plane with two different elements, the material world is three-dimensional and consists of almost 100 different atoms and millions of molecules, each with a mass, a charge, a shape and properties that give rise to special properties , affinities and behaviors unique to each combination.
GNoME AI to save chemistry
Stable components make up the majority of our knowledge.
For example, sodium chloride (NaCl salt) is a compound that crystallizes easily and also dissolves easily. It is a common connection in our conditions. Compounds as simple as carbon can be organized in many different ways to yield materials as common as graphite, as rare as diamond, or as sophisticated as graphene or fullerenes. Complex combinations, for example based on cement (silica, alumina, iron oxide and lime), all consist of stable molecules, but their interactions are crucial for the properties of the end product.
In short, of the billions and billions of ways to connect atoms and molecules, only a certain number can form stable compounds and even fewer of them have special properties such as conductivity, elasticity, magnetism, strength, resistance, etc.
The challenge in materials research is to first determine whether a combination of atoms is stable and then try to predict its properties.
Based on data on nearly 40,000 known materials and existing theories such as density functional theory (DFT), GNoME, an artificial intelligence developed by DeepMind and also behind AlphaGo and AlphaFold, the foundation of protein data, was trained and then modeled more than 2 .2 million potential crystals, which would have taken around 600 years at the current rate of research.
Of those 2.2 million crystals, this AI identified that approximately 421,000 would be stable enough to be synthesized, and 738 of them were actually synthesized with properties consistent with predictions of around 80%. Compared to 33% in the best cases so far or 1% not so long ago, this is outrageous and opens up destabilizing prospects. AI does not yet take into account the impact of these hypothetical compounds on the environment, health or life. Will it possibly provide formulas for their recycling or recovery?
Currently, AI is limited to compounds with no more than 5 atom types and strategies for substitution, estimation of binding energy levels and structural forms. As learning progresses, more data will be collected and other, bolder approaches will radically increase the number of possibilities. Enough to keep the entire industry busy for many years.
Return of the Go player
In the history of AI and the game of Go, Kellin Pelrine, a graduate student in machine learning at McGill University, beat AlphaGo soundly 14 to 1 in 2023 by simply studying how AlphaGo worked and using a confounding strategy (recommended by another AI). , a strategy that any person would have recognized immediately.
GNoME's limitations are of the same nature and its potential only improves through human intervention, giving it theories, data and goals to pursue. It is an extraordinary tool whose limitations 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 lost the game to a human (assisted by an AI)
https://www.20minutes.fr/high-tech/4024717-20230221-intelligence-artificielle-perdu-jeu-go-face-homme-aide-ia
Art of Go – Rules of the game of Go – https://artdugo.fr/regles-du-jeu-de-go/
Find out more about this technology