Watch 44 million atoms simulated using AI and a supercomputer

The most accurate simulation of objects made from tens of millions of atoms has been run on one the world’s top supercomputers with the help of artificial intelligence.

Existing simulations that describe in detail how atoms behave, interact and evolve are limited to small molecules, because of the computational power needed. There are techniques to simulate much larger numbers of atoms through time, but these rely on approximations and aren’t accurate enough to extract many detailed features of the molecule in question.

Now, Boris Kozinsky at Harvard University and his colleagues have developed a tool, called Allegro, that can accurately simulate systems with tens of millions of atoms using artificial intelligence.

Kozinsky and his team used the world’s 8th most powerful supercomputer, Perlmutter, to simulate the 44 million atoms involved in the protein shell of HIV. They also simulated other common biological molecules such as cellulose, a protein missing in people with haemophilia and a widespread tobacco plant virus.

“Anything that’s essentially made out of atoms, you can simulate with these methods at extremely high accuracy, and now also at large scale,” says Kozinsky. “This is one demonstration, but by no means constrained to this domain.” The system could also be used for many problems in materials science, such as investigating batteries, catalysis and semiconductors, he says.

To be able to simulate such large numbers of particles, the researchers used a kind of AI called a neural network to calculate interactions between atoms that were symmetrical from every angle, a principle called equivariance.

“When you develop networks that very fundamentally include these symmetries… you get these big improvements in accuracy and other properties that we care about, such as the stability of simulations, or how fast the machine learning model learns as you teach it with more data,” says team member Albert Musaelian, also at Harvard.

“This is a tour de force in programming and demonstrating that these machine-learned potentials are now scalable,” says Gábor Csányi at the University of Cambridge.

However, simulating biological molecules like these is more of a demonstration that the tool works for large systems than a practical boost for researchers, as biochemists already have accurate enough tools that can be run much faster, he says. Where it could be useful is for materials with lots of atoms that experience shocks and extreme forces over very short timescales, such as in planetary cores, says Csányi.