Artificial intelligence has been used to make simulations that track billions of interacting molecules at unprecedented scale and speed, paving the way for a greater understanding of material science, physics and biology.
When it comes to simulating the interaction of groups of atoms or molecules, there are two schools of thought. One approach, known as “ab initio”, strictly models all the Newtonian and quantum forces at play. The other uses various approximations to reduce the enormous computational burden. The precision of the former is key to some work, particularly in physics, while close approximations are adequate for engineering and biology.
The current world record for approximate modelling is a simulation of 100 billion atoms, while ab initio simulations have so far been limited to 100 million atoms using the Summit supercomputer at Oak Ridge National Laboratory in Tennessee.
Advertisement
Until June 2020, Summit was the fastest supercomputer in the world, but a full day of computation still only managed to simulate less than 2 nanoseconds of atomic or molecular interactions. Now, an international group of researchers has optimised code for ab initio systems to reduce demands on memory and processors, retaining almost all accuracy while drastically reducing the size and speed of computer required. Using the same Summit supercomputer as in previous work, the team accelerated simulations by up to seven times, churning through 11.2 nanoseconds of simulation time each day.
The researchers took a series of interactions between small numbers of atoms and worked out the precise forces at play, then used this data to train an AI to approximate similar problems quickly and achieve good accuracy without extensive computation. This let them improve the speed and also scale up to larger systems, creating a model��with 3.4 billion copper atoms.
They also ran the same model on what is currently the world’s most powerful supercomputer, Fugaku, at the RIKEN Center for Computational Science in Japan. This could run the model 21.2 times faster than Summit for water molecules and 46.7 times faster for copper molecules.
Although they weren’t able to access the whole of the Fugaku machine for experiments at once, they used smaller tests – which are common in supercomputing due to the expense and demand on high-power machines – to prove that the computer is capable of running models of up to 17 billion atoms of copper and 25 billions atoms of water.
While 25 billion water molecules is certainly nothing to be sniffed at, the researchers would need to scale that up 60 billion times to model the approximately 1.5 sextillion molecules in just a single drop of water. This would be far beyond the reach of even the most powerful computers yet built.
at Imperial College London says the result of the work is a system with accuracy close to that of ab initio simulations, but with computational requirements closer to simulations that use approximations – potentially the best of both worlds.
“They’re taking a very accurate calculation of the energies in the system, and through machine learning they’re converting these extremely complex calculations into something which is… numerically much easier to calculate,” says Muller.
“It’s unprecedented numbers considering the accuracy, and that’s the key point. They’re stretching the frontiers of computing and have found a real use for one of these billion-dollar machines,” he says.
The researchers say in their paper that the model opens the door to studying the mechanical properties of metals, semiconductors or batteries in a more realistic scale and detail.
Reference:��
