DeepMind logo.
Dec. 10, 2021
Machine-learning algorithm predicts material properties using electron density.
A team led by scientists at the London-based artificial-intelligence company DeepMind has developed a machine-learning model that suggests a molecule’s characteristics by predicting the distribution of electrons within it. The approach, described in the 10 December issue of Science (1), can calculate the properties of some molecules more accurately than existing techniques.
Image above: The AI predicts the distribution of electrons within a molecule (illustration) and uses it to calculate physical properties. Image Credit: DeepMind.
“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 paper is “a solid piece of work”, says Katarzyna Pernal, a computational chemist at Lodz University of Technology in Poland. But she adds that the machine-learning model has a long way to go before it can be useful for computational chemists.
Predicting properties
In principle, the structure of materials and molecules is entirely determined by quantum mechanics, and specifically by the Schrödinger equation, which governs the behaviour of electron wavefunctions. These are the mathematical gadgets that describe the probability of finding a particular electron at a particular position in space. But because all the electrons interact with one another, calculating the structure or molecular orbitals from such first principles is a computational nightmare, and can be done only for the simplest molecules, such as benzene, says James Kirkpatrick, a physicist at DeepMind.
To get around this problem, researchers — from pharmacologists to battery engineers — whose work relies on discovering or developing new molecules have for decades relied on a set of techniques called density functional theory (DFT) to predict molecules’ physical properties. The theory does not attempt to model individual electrons, but instead aims to calculate the overall distribution of the electrons’ negative electric charge across the molecule. “DFT looks at the average charge density, so it doesn’t know what individual electrons are,” says Kirkpatrick. Most properties of matter can then be easily calculated from that density.
Since its beginnings in the 1960s, DFT has become one of the most widely used techniques in the physical sciences: an investigation by Nature’s news team in 2014 found that, of the top 100 most-cited papers, 12 were about DFT. Modern databases of materials’ properties, such as the Materials Project, consist to a large extent of DFT calculations.
But the approach has limitations, and is known to give the wrong results for certain types of molecule, even some as simple as sodium chloride. And although DFT calculations are vastly more efficient than those that start from basic quantum theory, they are still cumbersome and often require supercomputers. So, in the past decade, theoretical chemists have increasingly started to experiment with machine learning, in particular to study properties such as materials’ chemical reactivity or their ability to conduct heat.
Ideal problem
The DeepMind team has made probably the most ambitious attempt yet to deploy AI to calculate electron density, the end result of DFT calculations. “It’s sort of the ideal problem for machine learning: you know the answer, but not the formula you want to apply,” says Aron Cohen, a theoretical chemist who has long worked on DFT and who is now at DeepMind.
The team trained an artificial neural network on data from 1,161 accurate solutions derived from the Schrödinger equations. To improve accuracy, they also hard-wired some of the known laws of physics into the network. They then tested the trained system on a set of molecules that are often used as a benchmark for DFT, and the results were impressive, says von Lilienfeld. “This is the best the community has managed to come up with, and they beat it by a margin,” he says.
One advantage of machine learning, von Lilienfeld adds, is that although it takes a massive amount of computing power to train the models, that process needs to be done only once. Individual predictions can then be done on a regular laptop, vastly reducing their cost and carbon footprint, compared with having to perform the calculations from scratch every time.
Kirkpatrick and Cohen say that DeepMind is releasing their trained system for anyone to use. For now, the model applies mostly to molecules and not to the crystal structures of materials, but future versions could work for materials, too, the authors say.
doi: https://doi.org/10.1038/d41586-021-03697-8
Reference:
1. Kirkpatrick, J. et al. Science 374, 1385–1389 (2021). https://doi.org/10.1126%2Fscience.abj6511
Related article:
DeepMind’s AI helps untangle the mathematics of knots
https://orbiterchspacenews.blogspot.com/2021/12/deepminds-ai-helps-untangle-mathematics.html
Image (mentioned), Text, Credits: Nature/Davide Castelvecchi.
Greetings, Orbiter.ch
