AI Accelerates Simulations of Irregular Particles

Simulating particles is relatively easy when they are perfectly spherical, but real-world particles are often irregular in shape and size, making simulations much more complex and resource-intensive.

Understanding particle behavior is crucial, particularly with the increasing presence of microplastics—tiny fragments of plastic that have become widespread environmental pollutants. These microplastics result from the breakdown of larger plastic waste through mechanical forces or UV degradation, and they are now found almost everywhere. To address this environmental challenge, it's vital to better understand the behavior of these non-uniform particles.

Researchers at the University of Illinois Urbana-Champaign have developed a novel approach using neural networks to predict how irregularly shaped particles interact, which accelerates molecular dynamics simulations. This method makes simulations up to 23 times faster compared to traditional techniques and is adaptable to various shapes if provided with adequate training data.

Their study, titled "Molecular Dynamics Simulations of Anisotropic Particles Accelerated by Neural-Net Predicted Interactions," appears in The Journal of Chemical Physics.

Antonia Statt, a professor of materials science and engineering, points out that simulating irregular shapes is more challenging than spherical ones. Unlike spheres, which only require calculating the distance between their centers, irregular shapes like cubes or cylinders require additional details about their angles and relative positions. Traditional methods involve approximating complex shapes with numerous small spheres, a process that is both inefficient and costly.

Statt's team has employed machine learning techniques, specifically a feed-forward neural network, to streamline this process. By training the neural network with sufficient data, researchers can bypass the need to calculate every interaction between individual spheres, resulting in faster simulations that are as accurate as conventional methods.

In the future, Statt aims to extend this technique to handle even more complex and mixed shapes, such as combinations of cubes and cylinders. The versatility of this method could enable the simulation of a wide variety of particle interactions.

Additional contributors to this research include B. Ruşen Argun from the Department of Mechanical Engineering and Yu Fu from the Department of Physics at Illinois.