AI recreates the periodic table of elements of chemistry

It took nearly a century of trial and error for human scientists to organize the periodic table of elements, arguably one of the greatest scientific achievements in chemistry, in its history. current form.

A new artificial intelligence (AI) program developed by Stanford physicists accomplished the same feat in just a few hours.

Called Atom2Vec, the program has learned to distinguish different atoms after analyzing a list of chemical compound names from an online database. The unsupervised AI then used concepts borrowed from the field of natural language processing – in particular, the idea that the properties of words can be understood by looking at the other words that surround them – to group the elements together. according to their chemical properties.

"We wanted to know if an AI could be smart enough to discover the periodic table alone, and our team showed that she could do it," said Shou-Cheng Zhang, JG Jackson and CJ Wood's professor of physics at Stanford School. human sciences and sciences.

Zhang says that the research, published in the June 25 issue Proceedings of the National Academy of Sciences, is an important first step towards a more ambitious goal, which is to design a replacement of the Turing Test – the current gold standard for the intelligence of the gauges.

For an AI to pass the Turing test, she must be able to answer questions written in an indistinguishable manner from a human. But Zhang thinks the test is flawed because it's subjective. "Humans are the product of evolution and our minds are cluttered with all sorts of irrationalities. For an AI to pass the Turing test, it would need to replicate all our human irrationalities," Zhang said. . "It's very hard to do, and it's not a good use of programmers' time."

Zhang would rather propose a new reference in artificial intelligence. "We want to see if we can design an AI capable of beating humans by discovering a new law of nature," he said. "But to do this, we must first test if our AI can do some of the greatest discoveries already made by humans."

By recreating the periodic table of elements, Atom2Vec achieved this secondary goal, Zhang said.

Potassium is king as …

Zhang and his group modeled Atom2Vec on an AI program that Google engineers created to analyze natural language. Called Word2Vec, the AI ​​language works by converting words into numeric codes, or into vectors. In analyzing vectors, the IA can estimate the probability of occurrence of a word in a text given the co-occurrence of other words.

For example, the word "king" is often accompanied by "queen" and "man" of "woman". Thus, the mathematical vector of the "king" could be roughly translated as "king = a queen minus a woman plus a man".

"We can apply the same idea to atoms," Zhang said. "Instead of feeding all the words and phrases of a collection of texts, we have powered Atom2Vec with all known chemical compounds, such as NaCl, KCl, H20, and so on."

From these scattered data, the AI ​​program included, for example, that potassium (K) and sodium (Na) should have similar properties because the two elements can bind chlorine (Cl). "Just as the king and queen are similar, potassium and sodium are similar," Zhang said.

Zhang hopes that in the future, scientists will be able to harness Atom2Vec's knowledge to discover and design new materials. "For this project, the AI ​​program was not supervised, but you can imagine giving it a goal and asking it to find, for example, a very effective material for converting sunlight into energy "Zhang said.

His team is already working on version 2.0 of its AI program, which will focus on solving an insoluble problem in medical research: designing antibodies that can attack antigens – molecules capable of breaking down. Induce an immune response – specific to cancer cells. Currently, one of the most promising approaches to cure cancer is cancer immunotherapy, which involves harnessing antibodies that can attack antigens on cancer cells.

But the human body can produce more than 10 million unique antibodies, each of which consists of a different combination of about 50 genes. "If we can map these building block genes to a mathematical vector, then we can organize all the antibodies into something similar to a periodic table," says Zhang. "Then, if you discover that an antibody is effective against an antigen but that it is toxic, you can search in the same family for another antibody that is as effective but less toxic."

Explore further:
Speakers store abstract information, whatever their language

More information:
Quan Zhou el al., "Atom2Vec: Learning atoms for the discovery of materials" PNAS (2018). www.pnas.org/cgi/doi/10.1073/pnas.1801181115