A team led by scientists from the London-based artificial intelligence company DeepMind has developed a machine learning model that suggests the properties of a molecule by predicting the distribution of electrons in it. The approach described in the December 10th issue of Science1 can calculate the properties of a few more molecules more accurate than existing techniques.
“To make it as accurate as they have done is a feat,” says Anatole von Lilienfeld, a materials scientist at the University of Vienna.
The article is “a solid piece of work,” says Katarzyna Pernal, computer chemist at the Technical University of Lodz in Poland. However, he adds that the machine learning model has a long way to go before it can be useful to computational chemists.
In principle, the structure of materials and molecules is completely determined by quantum mechanics, namely by the Schrödinger equation, which determines the behavior of the wave functions of electrons. But because all the electrons interact with each other, calculating the molecular structure, or orbitals, from these first principles is a computational nightmare and can only be done for the simplest of molecules like benzene, he says. James Kirkpatrick, physicist at DeepMind.
To solve this problem, researchers, from pharmacologists to battery engineers whose work is based on the discovery or development of new molecules, have for decades relied on a number of techniques known as density functional theory (DFT) to understand the physical properties of Predict molecules. The theory does not try to model individual electrons, but tries to calculate the general distribution of the negative electric charge of electrons in the entire molecule. “DFT examines the average charge density, so it doesn’t know what individual electrons are,” says Kirkpatrick. Most of the properties of matter can easily be calculated from this density.
Since its inception in the 1960s, DFT has become one of the most widely used techniques in the natural sciences: Research by the Nature News team in 2014 found that of the 100 most cited articles, 12 revolved around DFT. Properties like the Materials Project consist largely of DFT calculations.
However, the approach has limitations and is known to produce incorrect results for certain types of molecules, even some as simple as sodium chloride. And although DFT calculations are much more efficient than those based on basic quantum theory, they are still cumbersome and often require supercomputers. For this reason, theoretical chemists have increasingly begun to experiment with machine learning over the past decade, particularly to study properties such as the chemical reactivity of materials or their ability to conduct heat.
Ideal problem
The DeepMind team has made what is possibly the most ambitious attempt to implement AI to compute electron density, the end result of DFT calculations. It’s kind of an ideal machine learning problem – you know the answer, but you don’t know the formula you’re trying to use, says Aron Cohen, longtime theoretical chemist at DFT and now at DeepMind.
The team trained an artificial neural network with data from 1,161 precise solutions derived from the Schrödinger equations. To improve accuracy, they also linked some of the well-known laws of physics to the internet. They then tested the trained system on a number of molecules that are often used as benchmarks for DFT, and the results were impressive, says von Lilienfeld. This is the best the community has ever achieved, and they far exceeded it, he says.
One advantage of machine learning, von Lilienfeld adds, is that training the models requires a lot of computing power, but only needs to be carried out once. Individual predictions can be made on a normal laptop, which significantly reduces costs. and carbon footprint, compared to having to do the calculations from scratch every time.
Kirkpatrick and Cohen say DeepMind is sharing their trained system with everyone. Currently, the model applies mostly to molecules rather than the crystalline structure of materials, but future versions could work for materials as well, the authors say.
