DeepMind AI: Machine learning tool helps study strange electrons in chemical reactions

Strange so-called fractional electrons are crucial to many chemical reactions, but traditional methods cannot model them – a problem that DeepMind has used machine learning to fix


9 December 2021

An artistic representation of electrons

An artistic representation of molecules interacting


Machine-learning tools have taken us closer to understanding electrons and how they behave in chemical interactions, following news that UK-based AI company DeepMind, owned by Google’s parent company Alphabet, has created a tool that solves a fundamental problem with how we model chemistry.

The tool, called DeepMind 21, is based on a modelling method called density functional theory (DFT), which relates the location of electrons in a given group of atoms to the total energy the atoms share to determine the chemical and physical properties of a molecule or material. “DFT is a very widely used tool and it’s usually very effective, but it has these failures, so tracking down and understanding these failures is important,” says DeepMind’s Aron Cohen.

One of those failures is an inability to deal with fractional electrons, a theoretical construct in which the charge of an electron is split into multiple particles. Traditional DFT tools can model systems with one or two electrons, but they fail at modelling those with, say, 1.5 electrons, which is important in cases where an electron is shared between more than one atom.

“On the one hand, fractional electrons are fictitious objects, there’s no such thing as a fractional electron – electrons are whole by definition,” says James Kirkpatrick at DeepMind. “But by fixing these fractional electron problems, we are able to correctly describe chemical systems which usually have got these fundamental errors in their descriptions.”

DeepMind 21 works using machine learning, a process by which an artificial intelligence is fed a training set of data that includes both the relevant problems and their solutions. Through examining the training set, the AI learns to look for patterns and apply them to similar, incomplete data sets.

The researchers trained their AI with 2235 examples of chemical reactions, complete with information on the electrons involved and the energies of the systems. Of these, 1074 represented systems where fractional electrons would pose a problem to traditional DFT analyses.

Then, they applied the AI to chemical reactions that weren’t included in the training data. Not only did DeepMind 21 represent the fractional electrons correctly, but its results were more precise than traditional DFT analyses. It even worked on data about atoms with strange properties that didn’t closely resemble anything in the training data. While there are other methods that can create these models, they take far more computing power and time, says John Perdew at Temple University in Pennsylvania.

This is a major advance in terms of using machine learning to understand chemistry, says Perdew. “It suggests a unification of standard theoretical approaches, such as the satisfaction of exact theorems, with data-driven machine learning, a unification that may be more powerful than either approach by itself,” he says.

DeepMind has also announced that the AI’s code will be made open source, so chemists and materials researchers around the world will be able to apply it to a variety of problems. Fractional electrons are particularly relevant in organic chemistry, says Cohen, so it may be particularly useful in that field.

Journal reference: Science, DOI: 10.1126/science.abj6511

More on these topics: