DeepMind has created the best chess player, one of the most accurate weather forecasting systems and a poker player that could give the card sharks of Las Vegas a run for their money.
In its latest venture, the London-based artificial intelligence company has opened a window into an altogether stranger realm — quantum chemistry. Its researchers have built an AI system that can predict the chemical energy stored in a molecule, a property dictated by its structure that will decide how the group of atoms will behave in a certain environment and whether a particular reaction can occur.
The breakthrough opens the way for scientists seeking new drugs and materials to obtain more accurate results when they simulate chemical experiments on powerful computers.
The work, published in the journal Science, represents a significant advance in predicting how matter behaves at the nanoscale — the level of sub-atomic particles such as neutrons and electrons.
“Solving some of the major challenges of the 21st century, such as producing clean electricity or developing high temperature superconductors, will require us to design new materials with specific properties,” the research team said. “To do this on a computer requires the simulation of electrons, the subatomic particles that govern how atoms bond to form molecules and are also responsible for the flow of electricity in solids.”
Electrons are often represented in textbooks as whirling round the atom’s nucleus in neat elliptical orbits. But they are better thought of as clustering in swarms. Quantum scientists look at the probability of whether an electron will be found in a particular place at a certain time.
Roughly speaking, it is possible to build a kind of three-dimensional heat map that tells you where electrons are most likely to be, known as the electron density. A breakthrough came in the 1960s, when it was realised that if the electron density were known, a formula could be used to determine the energy of a molecule.
This insight gave rise to the dominant method of predicting the properties of electron interactions in chemistry, biology and materials science. Walter Kohn, the theoretical chemist, won a Nobel prize in 1998 for his work in this field.
The snag is that after more than 50 years nobody has been able to work out what the formula, known as the density functional, is exactly.
DeepMind, which is owned by Google’s parent company Alphabet, approached the problem by training a computer system, known as a neural network.
Essentially, the computer was shown the electron density of molecules, whose energies were known.
The AI was able, in essence, to provide an improved approximation of the density functional. It could then make improved predictions of how molecules behaved, resulting in a better description of a range of chemical reactions.
Ultimately, it is hoped that this will lead to the creation of useful new molecules, which could be used for novel drugs and other materials. “The quantum world is a little like a dark room,” James Kirkpatrick, a research scientist at DeepMind, said.
“At the moment, people proceed by trial and error. These computational tools, they are a torch that you can shine and predict what is around you.”
DeepMind delves to the depths of life
•In 2016 a DeepMind computer program called AlphaGo beat Lee Sedol, one of the world’s top players of the ancient game of Go. The 37th move in the second of five games left aficionados stunned. “The Google machine made a move that no human ever would,” one pundit said. “And it was beautiful.”
•A year later a more general program, AlphaZero, learnt to become the most powerful computer system to play chess and shogi (Japanese chess) by competing against itself for a few days.
•Having cut its teeth on games, DeepMind stopped playing around this year when it unveiled AlphaFold. The computer program achieved one of biology’s grand challenges — how to predict the structure of a protein from the genetic sequence that encodes it. The breakthrough promises to accelerate research in the life sciences and the discovery of new drugs. In a single day DeepMind released information on the shapes of 350,000 proteins, the building blocks of life, more than doubling the number understood by science.
•What’s next? Demis Hassabis, the DeepMind chief executive, told The Times this year that his team was working on nuclear fusion, potentially a source of essentially limitless energy, as well as quantum chemistry and weather prediction.
“When I say that out loud, it sounds crazy,” he said. “But we’re actually working on all these things. And we usually complete the job when we start working on something.”
Since the interview, in July, DeepMind has revealed advances in quantum chemistry and predicting the weather.