Using machine learning to accelerate simulations of irregularly shaped particles

Simulating particles is a relatively simple task when those particles are spherical. In the real world, however, most particles are not perfect spheres, but take on irregular and variable shapes and sizes. Simulating these particles becomes a much more difficult and time-consuming task.

The ability to simulate particles is essential to understanding their behavior. For example, microplastics are a new form of pollution, as plastic waste has increased drastically and is degrading uncontrollably in the environment, either by mechanical means or through UV degradation. These tiny particles are now found almost everywhere in the world. To be able to address this environmental crisis, it is important to better understand these particles and their behavior.

To address this challenge, researchers at the University of Illinois at Urbana-Champaign trained neural networks to predict interactions between irregularly shaped particles to speed up molecular dynamics simulations. Using this method, simulations can be run up to 23 times faster than traditional simulation methods and can be applied to any irregular shape with sufficient training data.

This research, entitled “Molecular dynamics simulations of anisotropic particles accelerated by interactions predicted by neural networks”, is published in Journal of Chemical Physics.

“Microplastics are now everywhere in the environment and most of them are not spheres, they are very heterogeneous and have corners and edges. To solve the problem of their behavior in the environment, we need to develop new methods, find ways to simulate them faster, cheaper and more efficiently,” says Antonia Statt, professor of materials science and engineering.

Spheres are easy to simulate because the only parameter needed to determine how two particles interact is the distance between the centers of the spheres. To go from a sphere to more complex shapes, such as cubes or cylinders, you need to know not only the distance between two particles, but also the angles and relative positions of each particle. The traditional way to simulate cubes, for example, is to build the cube from several smaller spheres.

“It’s a very roundabout way of describing a cube, of tessellated it with little spheres,” Statt says. “It’s also expensive because you have to calculate the interactions of all the little spheres with each other. To get around that, we used machine learning, a feedforward neural network, which is a fancy way of saying, ‘Let’s adapt to a complicated function that we don’t know.’ And neural networks are really good at that. If you give them enough data, they can adapt to anything you want.”

With this method, it is not necessary to calculate all the distances between the small spheres individually. Only the center-to-center distance of the cube and its relative orientation are needed, which makes the calculation much easier and faster. In addition, this method is as accurate as traditional methods. It cannot be more accurate since it is formed from data produced from traditional methods, but it can be more efficient.

In the future, Statt would like to be able to simulate more complex irregular shapes as well as mixtures of different shapes, such as a cube and a cylinder, rather than two cubes. “We’ll have to learn all the individual interactions, but the method is general enough that we can do that,” she says.

Other contributors to this work include B. Ruşen Argun (Department of Mechanical Engineering, Illinois) and Yu Fu (Department of Physics, Illinois).